From 162db479c64de43fb77bdca2c6a8999a8a6643a4 Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Thu, 10 Sep 2026 13:04:05 +0200
Subject: [PATCH 01/13] feat(nebul): add Nebul provider with 16 models and
 hourly sync module

---
 .../alibaba/qwen3-30b-a3b-instruct-2507.toml  |  28 +++
 .../ministral-3-14b-instruct-2512.toml        |  19 ++
 packages/core/src/sync/index.ts               |   5 +-
 packages/core/src/sync/providers/nebul.ts     | 167 ++++++++++++++++++
 packages/core/test/nebul.test.ts              | 106 +++++++++++
 providers/nebul/logo.svg                      |   5 +
 .../Qwen/Qwen3-30B-A3B-Instruct-2507.toml     |   6 +
 .../nebul/models/Qwen/Qwen3.5-397B-A17B.toml  |  13 ++
 .../nebul/models/Qwen/Qwen3.8-27B-FP8.toml    |  14 ++
 .../deepseek-ai/DeepSeek-V4.1-Flash.toml      |  13 ++
 .../nebul/models/google/gemma-4-31B-it.toml   |   8 +
 .../models/meta-models/muse-glimmer-30b.toml  |  11 ++
 .../Ministral-3-14B-Instruct-2512.toml        |   5 +
 .../Mistral-Large-3-675B-Instruct-2512.toml   |   6 +
 .../mistralai/Mistral-Medium-3.5-128B.toml    |  10 ++
 .../nebul/models/moonshotai/Kimi-K3.toml      |  14 ++
 .../NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml  |  10 ++
 ...VIDIA-Nemotron-3-Super-120B-A12B-BF16.toml |  13 ++
 ...ron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml |  13 ++
 .../nebul/models/openai/gpt-oss-120b.toml     |  11 ++
 .../nebul/models/zai-org/GLM-5.3-Flash.toml   |  13 ++
 providers/nebul/models/zai-org/GLM-5.3.toml   |  14 ++
 providers/nebul/provider.toml                 |   5 +
 23 files changed, 508 insertions(+), 1 deletion(-)
 create mode 100644 models/alibaba/qwen3-30b-a3b-instruct-2507.toml
 create mode 100644 models/mistral/ministral-3-14b-instruct-2512.toml
 create mode 100644 packages/core/src/sync/providers/nebul.ts
 create mode 100644 packages/core/test/nebul.test.ts
 create mode 100644 providers/nebul/logo.svg
 create mode 100644 providers/nebul/models/Qwen/Qwen3-30B-A3B-Instruct-2507.toml
 create mode 100644 providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
 create mode 100644 providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
 create mode 100644 providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
 create mode 100644 providers/nebul/models/google/gemma-4-31B-it.toml
 create mode 100644 providers/nebul/models/meta-models/muse-glimmer-30b.toml
 create mode 100644 providers/nebul/models/mistralai/Ministral-3-14B-Instruct-2512.toml
 create mode 100644 providers/nebul/models/mistralai/Mistral-Large-3-675B-Instruct-2512.toml
 create mode 100644 providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
 create mode 100644 providers/nebul/models/moonshotai/Kimi-K3.toml
 create mode 100644 providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
 create mode 100644 providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
 create mode 100644 providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
 create mode 100644 providers/nebul/models/openai/gpt-oss-120b.toml
 create mode 100644 providers/nebul/models/zai-org/GLM-5.3-Flash.toml
 create mode 100644 providers/nebul/models/zai-org/GLM-5.3.toml
 create mode 100644 providers/nebul/provider.toml

diff --git a/models/alibaba/qwen3-30b-a3b-instruct-2507.toml b/models/alibaba/qwen3-30b-a3b-instruct-2507.toml
new file mode 100644
index 00000000000..94b773e0da2
--- /dev/null
+++ b/models/alibaba/qwen3-30b-a3b-instruct-2507.toml
@@ -0,0 +1,28 @@
+# Sources (accessed 2026-09-10):
+# https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507
+# Hub createdAt 2025-07-28 is the release; lastModified 2025-09-17 is the latest revision.
+
+name = "Qwen3 30B A3B Instruct 2507"
+description = "Updated non-thinking Qwen3 MoE with long context for instruction following and tool use"
+family = "qwen"
+release_date = "2025-07-28"
+last_updated = "2025-09-17"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+license = "Apache 2.0"
+
+[limit]
+context = 262_144
+output = 16_384
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507"
diff --git a/models/mistral/ministral-3-14b-instruct-2512.toml b/models/mistral/ministral-3-14b-instruct-2512.toml
new file mode 100644
index 00000000000..7cd61ebb5a3
--- /dev/null
+++ b/models/mistral/ministral-3-14b-instruct-2512.toml
@@ -0,0 +1,19 @@
+name = "Ministral 3 14B Instruct"
+description = "Efficient Mistral edge model balancing speed and cost for general-purpose tasks"
+family = "ministral"
+release_date = "2025-12-02"
+last_updated = "2025-12-02"
+attachment = true
+reasoning = false
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 262_144
+output = 262_144
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
diff --git a/packages/core/src/sync/index.ts b/packages/core/src/sync/index.ts
index 7f88aacb5bb..22a37b91b10 100644
--- a/packages/core/src/sync/index.ts
+++ b/packages/core/src/sync/index.ts
@@ -27,6 +27,7 @@ import { kilo } from "./providers/kilo.js";
 import { llmgateway, llmgatewayProviders } from "./providers/llmgateway.js";
 import { mergeGateway } from "./providers/merge-gateway.js";
 import { meta } from "./providers/meta.js";
+import { nebul } from "./providers/nebul.js";
 import { nanoGpt } from "./providers/nano-gpt.js";
 import { openai } from "./providers/openai.js";
 import { ofox } from "./providers/ofox.js";
@@ -152,6 +153,7 @@ export const providers: {
   "llmgateway-providers": SyncProvider<any>;
   "merge-gateway": SyncProvider<any>;
   meta: SyncProvider<any>;
+  nebul: SyncProvider<any>;
   "nano-gpt": SyncProvider<any>;
   ofox: SyncProvider<any>;
   openai: SyncProvider<any>;
@@ -187,6 +189,7 @@ export const providers: {
   "llmgateway-providers": llmgatewayProviders,
   "merge-gateway": mergeGateway,
   meta,
+  nebul,
   "nano-gpt": nanoGpt,
   ofox,
   openai,
@@ -219,7 +222,7 @@ export const groups = {
     "vercel",
   ],
   cloudflare: ["cloudflare-ai-gateway", "cloudflare-workers-ai"],
-  direct: ["ambient", "anthropic", "baseten", "chutes", "cortecs", "deepinfra", "digitalocean", "github-copilot", "google", "hyper", "meta", "openai", "ovhcloud", "pioneer", "tinfoil", "venice", "wandb", "xai"],
+  direct: ["ambient", "anthropic", "baseten", "chutes", "cortecs", "deepinfra", "digitalocean", "github-copilot", "google", "hyper", "meta", "nebul", "openai", "ovhcloud", "pioneer", "tinfoil", "venice", "wandb", "xai"],
 } as const;
 
 type ProviderID = keyof typeof providers;
diff --git a/packages/core/src/sync/providers/nebul.ts b/packages/core/src/sync/providers/nebul.ts
new file mode 100644
index 00000000000..ea29002e259
--- /dev/null
+++ b/packages/core/src/sync/providers/nebul.ts
@@ -0,0 +1,167 @@
+import { existsSync, readdirSync } from "node:fs";
+import path from "node:path";
+import { z } from "zod";
+
+import type { ExistingModel, SyncProvider, SyncedModel } from "../index.js";
+import { factorBaseModel } from "./openrouter.js";
+
+const API_ENDPOINT = "https://api.inference.nebul.io/model/info";
+const MODELS_DIR = path.join(import.meta.dirname, "..", "..", "..", "..", "..", "models");
+
+// Served org prefix -> models/ metadata namespace (HF org names differ from catalog labs).
+const ORG_TO_MODEL_PROVIDER: Record<string, string | undefined> = {
+  "deepseek-ai": "deepseek",
+  google: "google",
+  "meta-models": "meta",
+  mistralai: "mistral",
+  moonshotai: "moonshotai",
+  nvidia: "nvidia",
+  openai: "openai",
+  Qwen: "alibaba",
+  "zai-org": "zhipuai",
+};
+
+// Served IDs whose canonical metadata lives under a differently-named lab entry.
+const BASE_MODEL_ALIASES: Record<string, string | undefined> = {
+  "mistralai/Mistral-Large-3-675B-Instruct-2512": "mistral/mistral-large-2512",
+  "mistralai/Mistral-Medium-3.5-128B": "mistral/mistral-medium-2604",
+};
+
+// The synthetic health-check model and serving artifacts that must not enter the catalog.
+const PING_MODEL = "Nebul/Ping";
+const DENYLIST = /OCR|Qwen3Guard/i;
+
+// Deprecated server-side (descriptions point at GLM-5.3) but not flagged by /model/info.
+const DEPRECATED = new Set(["zai-org/GLM-5.1-FP8", "zai-org/GLM-5.2-FP8"]);
+
+const EffortValues = z.enum(["none", "minimal", "low", "medium", "high", "xhigh", "max", "default"]);
+
+const ModelInfo = z.object({
+  description: z.string().nullable().optional(),
+  huggingface_id: z.string().nullable().optional(),
+  input_cost_per_1m_tokens: z.number().nullable().optional(),
+  output_cost_per_1m_tokens: z.number().nullable().optional(),
+  cache_read_input_cost_per_1m_tokens: z.number().nullable().optional(),
+  max_input_tokens: z.number().nullable().optional(),
+  mode: z.string().nullable(),
+  model_type: z.string().nullable(),
+  reasoning_efforts: z.array(EffortValues).nullable().optional(),
+}).passthrough();
+
+export const NebulEntry = z.object({
+  model_info: ModelInfo,
+  model_name: z.string().min(1),
+}).passthrough();
+
+export const NebulResponse = z.object({
+  data: z.array(NebulEntry),
+}).passthrough();
+
+export type NebulEntry = z.infer<typeof NebulEntry>;
+
+export const nebul = {
+  id: "nebul",
+  name: "Nebul",
+  modelsDir: "providers/nebul/models",
+  async fetchModels() {
+    const response = await fetch(API_ENDPOINT);
+    if (!response.ok) {
+      throw new Error(`Nebul models request failed: ${response.status} ${response.statusText}`);
+    }
+    return response.json();
+  },
+  parseModels(raw) {
+    return NebulResponse.parse(raw).data;
+  },
+  // Unauthenticated /model/info is the authoritative catalog: entries removed
+  // server-side are removed here (deleteMissing defaults on), and new resolvable
+  // chat models are created with base_model overrides only.
+  translateModel(entry, context) {
+    if (!isCatalogChatModel(entry)) return undefined;
+    const id = entry.model_name;
+    const info = entry.model_info;
+    const baseModel = resolveBaseModel(id, info.huggingface_id ?? undefined);
+    if (baseModel === undefined) return undefined;
+    if (info.input_cost_per_1m_tokens == null || info.output_cost_per_1m_tokens == null || info.max_input_tokens == null) return undefined;
+    const values = {
+      reasoning_options: buildReasoningOptions(entry, context.existing(id)),
+      cost: {
+        input: info.input_cost_per_1m_tokens,
+        output: info.output_cost_per_1m_tokens,
+        cache_read: info.cache_read_input_cost_per_1m_tokens ?? undefined,
+      },
+      limit: { context: info.max_input_tokens },
+    };
+    return {
+      id,
+      model: factorBaseModel(baseModel, values, values.limit) as SyncedModel,
+    };
+  },
+  // Only report chat models whose base_model could not be resolved; filtered
+  // serving artifacts (embeddings, rerankers, the ping model) skip silently.
+  sourceID(entry: NebulEntry) {
+    return isCatalogChatModel(entry) ? entry.model_name : undefined;
+  },
+  skippedNotice(ids) {
+    if (ids.length === 0) return [];
+    return [
+      `Nebul models could not be resolved to lab metadata and need hand-authored base_model targets:`,
+      ids.map((id) => `\`${id}\``).join(", "),
+    ];
+  },
+} satisfies SyncProvider<NebulEntry>;
+
+function isCatalogChatModel(entry: NebulEntry): boolean {
+  const info = entry.model_info;
+  return info.model_type === "llm" && info.mode === "chat"
+    && entry.model_name !== PING_MODEL && !DEPRECATED.has(entry.model_name) && !DENYLIST.test(entry.model_name);
+}
+
+function buildReasoningOptions(entry: NebulEntry, existing: ExistingModel | undefined) {
+  const efforts = entry.model_info.reasoning_efforts ?? [];
+  if (efforts.length === 0) return existing?.reasoning_options;
+  return [{ type: "effort" as const, values: efforts }];
+}
+
+function resolveBaseModel(servedID: string, huggingfaceID: string | undefined): string | undefined {
+  return baseModelCandidates(servedID, huggingfaceID).find(canonicalExists);
+}
+
+// existsSync is case-insensitive on Windows/macOS; verify the real on-disk filename case
+// so the resolved base_model matches the canonical metadata exactly (and CI on Linux).
+function canonicalExists(candidate: string): boolean {
+  const file = path.join(MODELS_DIR, `${candidate}.toml`);
+  if (!existsSync(file)) return false;
+  try {
+    return readdirSync(path.dirname(file)).includes(path.basename(file));
+  } catch {
+    return false;
+  }
+}
+
+function baseModelCandidates(servedID: string, huggingfaceID: string | undefined): string[] {
+  const alias = BASE_MODEL_ALIASES[servedID];
+  const servedCandidate = mapOrgToCandidate(servedID);
+  const hfCandidate = huggingfaceID === undefined ? undefined : mapOrgToCandidate(huggingfaceID);
+  return [
+    ...new Set([alias, servedCandidate, hfCandidate, ...quantizationStripped(hfCandidate), ...quantizationStripped(servedCandidate)]).values(),
+  ].filter((candidate): candidate is string => candidate !== undefined);
+}
+
+function mapOrgToCandidate(id: string): string | undefined {
+  const [org, ...modelParts] = id.split("/");
+  if (org === undefined || modelParts.length === 0) return undefined;
+  const provider = ORG_TO_MODEL_PROVIDER[org];
+  if (provider === undefined) return undefined;
+  return `${provider}/${modelParts.join("/").toLowerCase()}`;
+}
+
+// Hosts serve quantized checkpoints (e.g. -FP8, -BF16) of weights whose canonical
+// metadata is published for the base precision; try those names without the suffix.
+// NVIDIA also prefixes checkpoints with "NVIDIA-", which the metadata names drop.
+function quantizationStripped(candidate: string | undefined): string[] {
+  if (candidate === undefined) return [];
+  const withoutQuant = candidate.replace(/-(fp8|bf16|fp4|int8)$/i, "");
+  const withoutPrefix = withoutQuant.replace(/nvidia-/, "");
+  return withoutQuant === candidate ? [] : [...new Set([withoutQuant, withoutPrefix])].filter((value) => value !== candidate);
+}
diff --git a/packages/core/test/nebul.test.ts b/packages/core/test/nebul.test.ts
new file mode 100644
index 00000000000..e2d2aed1fc1
--- /dev/null
+++ b/packages/core/test/nebul.test.ts
@@ -0,0 +1,106 @@
+import { expect, test } from "bun:test";
+
+import type { ExistingModel } from "../src/sync/index.js";
+import {
+  NebulEntry,
+  NebulResponse,
+  nebul,
+} from "../src/sync/providers/nebul.js";
+
+function nebulEntry(model_name?: string, model_info: Record<string, unknown> = {}): NebulEntry {
+  return NebulEntry.parse({
+    model_name: model_name ?? "zai-org/GLM-5.3",
+    model_info: {
+      description: "test",
+      huggingface_id: "zai-org/GLM-5.3",
+      input_cost_per_1m_tokens: 1.47,
+      output_cost_per_1m_tokens: 4.62,
+      cache_read_input_cost_per_1m_tokens: 0.35,
+      max_input_tokens: 1_000_000,
+      mode: "chat",
+      model_type: "llm",
+      reasoning_efforts: ["low", "high", "max"],
+      ...model_info,
+    },
+  });
+}
+
+function existingWith(reasoning_options: ExistingModel["reasoning_options"]): ExistingModel {
+  return { reasoning_options } as ExistingModel;
+}
+
+const context = (existing: ExistingModel | undefined) => ({ existing: () => existing });
+
+test("syncs Nebul's factored overrides against resolved lab metadata", () => {
+  const translated = nebul.translateModel(nebulEntry("zai-org/GLM-5.3", { max_input_tokens: 1_048_576 }), context(undefined));
+  expect(translated).toMatchObject({
+    id: "zai-org/GLM-5.3",
+    model: {
+      base_model: "zhipuai/glm-5.3",
+      cost: { input: 1.47, output: 4.62, cache_read: 0.35 },
+      limit: { context: 1_048_576 },
+      reasoning_options: [{ type: "effort", values: ["low", "high", "max"] }],
+    },
+  });
+});
+
+test("preserves authored reasoning controls when the host exposes no efforts", () => {
+  const authored = [{ type: "toggle" as const }];
+  const translated = nebul.translateModel(nebulEntry("zai-org/GLM-5.3", { reasoning_efforts: [] }), context(existingWith(authored)));
+  expect(translated?.model.reasoning_options).toEqual(authored);
+});
+
+test("resolves base models across org renames and quantization suffixes", () => {
+  const cases: [string, string | null, string][] = [
+    ["Qwen/Qwen3.8-27B-FP8", "Qwen/Qwen3.8-27B-FP8", "alibaba/qwen3.8-27b"],
+    ["nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "nvidia/nemotron-3-nano-30b-a3b"],
+    ["mistralai/Mistral-Large-3-675B-Instruct-2512", "mistralai/Mistral-Large-3-675B-Instruct-2512", "mistral/mistral-large-2512"],
+    ["mistralai/Mistral-Medium-3.5-128B", "mistralai/Mistral-Medium-3.5-128B", "mistral/mistral-medium-2604"],
+  ];
+  for (const [model_name, huggingface_id, expected] of cases) {
+    const entry = nebulEntry(model_name, { huggingface_id });
+    expect(nebul.translateModel(entry, context(undefined))?.model.base_model).toBe(expected);
+  }
+});
+
+test("skips the ping model, serving artifacts, and deprecated entries silently", () => {
+  for (const model_name of ["Nebul/Ping", "zai-org/GLM-5.1-FP8", "zai-org/GLM-5.2-FP8", "Nebul-OCR/Some-OCR", "Qwen/Qwen3Guard-Something"]) {
+    const entry = nebulEntry(model_name, {});
+    expect(nebul.translateModel(entry, context(undefined))).toBeUndefined();
+    expect(nebul.sourceID(entry)).toBeUndefined();
+  }
+});
+
+test("skips embeddings and rerankers while reporting unresolvable chat models", () => {
+  const embedding = nebulEntry("BAAI/bge-m3", { model_type: "embedding" });
+  expect(nebul.translateModel(embedding, context(undefined))).toBeUndefined();
+  expect(nebul.sourceID(embedding)).toBeUndefined();
+
+  const chat = nebulEntry("mistralai/Mistral-Large-3-675B-Instruct-2512", { huggingface_id: null });
+  expect(nebul.translateModel(chat, context(undefined))).toBeDefined();
+  expect(nebul.sourceID(chat)).toBe("mistralai/Mistral-Large-3-675B-Instruct-2512");
+});
+
+test("skips chat models whose pricing or context is absent instead of crashing", () => {
+  const unpriced = nebulEntry("zai-org/GLM-5.3", { input_cost_per_1m_tokens: null, output_cost_per_1m_tokens: null, max_input_tokens: null });
+  expect(nebul.translateModel(unpriced, context(undefined))).toBeUndefined();
+  expect(nebul.sourceID(unpriced)).toBe("zai-org/GLM-5.3");
+});
+
+test("parses nullable serving artifacts and unknown-host metadata from /model/info", () => {
+  const parsed = NebulResponse.parse({
+    data: [
+      { model_name: "Some/Embedding", model_info: { mode: null, model_type: "embedding" } },
+      { model_name: "Some/Chat", model_info: { mode: "chat", model_type: "llm", unknown_host_field: true } },
+    ],
+  });
+  expect(parsed.data).toHaveLength(2);
+});
+
+test("rejects unknown reasoning effort values from the host", () => {
+  expect(() =>
+    NebulResponse.parse({
+      data: [{ model_name: "Some/Chat", model_info: { mode: "chat", model_type: "llm", reasoning_efforts: ["ultra"] } }],
+    }),
+  ).toThrow();
+});
diff --git a/providers/nebul/logo.svg b/providers/nebul/logo.svg
new file mode 100644
index 00000000000..33daff15792
--- /dev/null
+++ b/providers/nebul/logo.svg
@@ -0,0 +1,5 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor">
+  <g transform="translate(-70.45 1.54) scale(0.55)">
+    <path d="M159.479 5.82737V22.8078L144.803 8.13201C141.376 4.7047 135.818 4.7047 132.391 8.13201C130.677 9.84566 129.819 12.091 129.819 14.3383V33.3328H140.342V18.5582L155.117 33.3328H170L159.762 23.0944C165.442 22.9422 170 18.2954 170 12.5772V5.8412L159.477 5.82737H159.479Z"/>
+  </g>
+</svg>
diff --git a/providers/nebul/models/Qwen/Qwen3-30B-A3B-Instruct-2507.toml b/providers/nebul/models/Qwen/Qwen3-30B-A3B-Instruct-2507.toml
new file mode 100644
index 00000000000..2396ed198f9
--- /dev/null
+++ b/providers/nebul/models/Qwen/Qwen3-30B-A3B-Instruct-2507.toml
@@ -0,0 +1,6 @@
+base_model = "alibaba/qwen3-30b-a3b-instruct-2507"
+
+[cost]
+input = 0.21
+output = 0.74
+cache_read = 0.05
diff --git a/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml b/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
new file mode 100644
index 00000000000..b9cc203d0a9
--- /dev/null
+++ b/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
@@ -0,0 +1,13 @@
+base_model = "alibaba/qwen3.5-397b-a17b"
+
+# Controls per alibaba's official entry: thinking toggle + budget
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "budget_tokens"
+
+[cost]
+input = 0.63
+output = 3.78
+cache_read = 0.15
diff --git a/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
new file mode 100644
index 00000000000..e0d778d4b2a
--- /dev/null
+++ b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
@@ -0,0 +1,14 @@
+base_model = "alibaba/qwen3.8-27b"
+
+# Controls per deepinfra peer: thinking toggle + efforts
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "xhigh"]
+
+[cost]
+input = 0.21
+output = 0.74
+cache_read = 0.05
diff --git a/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml b/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
new file mode 100644
index 00000000000..d00053543cb
--- /dev/null
+++ b/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
@@ -0,0 +1,13 @@
+base_model = "deepseek/deepseek-v4.1-flash"
+
+# Efforts: reasoning_effort = low|high|max (from /model/info reasoning_efforts)
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
+[cost]
+input = 0.2
+output = 0.7
+
+[limit]
+context = 1_048_576
diff --git a/providers/nebul/models/google/gemma-4-31B-it.toml b/providers/nebul/models/google/gemma-4-31B-it.toml
new file mode 100644
index 00000000000..4c8e563df73
--- /dev/null
+++ b/providers/nebul/models/google/gemma-4-31B-it.toml
@@ -0,0 +1,8 @@
+base_model = "google/gemma-4-31b-it"
+
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0.42
+output = 1.05
diff --git a/providers/nebul/models/meta-models/muse-glimmer-30b.toml b/providers/nebul/models/meta-models/muse-glimmer-30b.toml
new file mode 100644
index 00000000000..0daac5a89cc
--- /dev/null
+++ b/providers/nebul/models/meta-models/muse-glimmer-30b.toml
@@ -0,0 +1,11 @@
+base_model = "meta/muse-glimmer-30b"
+
+# Efforts per fireworks peer (models without API-exposed levels keep lab controls)
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "xhigh"]
+
+[cost]
+input = 0.7
+output = 3.5
+cache_read = 0.35
diff --git a/providers/nebul/models/mistralai/Ministral-3-14B-Instruct-2512.toml b/providers/nebul/models/mistralai/Ministral-3-14B-Instruct-2512.toml
new file mode 100644
index 00000000000..d2768469116
--- /dev/null
+++ b/providers/nebul/models/mistralai/Ministral-3-14B-Instruct-2512.toml
@@ -0,0 +1,5 @@
+base_model = "mistral/ministral-3-14b-instruct-2512"
+
+[cost]
+input = 0.73
+output = 3.63
diff --git a/providers/nebul/models/mistralai/Mistral-Large-3-675B-Instruct-2512.toml b/providers/nebul/models/mistralai/Mistral-Large-3-675B-Instruct-2512.toml
new file mode 100644
index 00000000000..992a53758fd
--- /dev/null
+++ b/providers/nebul/models/mistralai/Mistral-Large-3-675B-Instruct-2512.toml
@@ -0,0 +1,6 @@
+base_model = "mistral/mistral-large-2512"
+
+[cost]
+input = 0.6
+output = 1.73
+cache_read = 0.15
diff --git a/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml b/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
new file mode 100644
index 00000000000..331f2d4f281
--- /dev/null
+++ b/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
@@ -0,0 +1,10 @@
+base_model = "mistral/mistral-medium-2604"
+
+# Efforts: reasoning_effort = low|medium|high|max (from /model/info reasoning_efforts)
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "max"]
+
+[cost]
+input = 1.65
+output = 8.25
diff --git a/providers/nebul/models/moonshotai/Kimi-K3.toml b/providers/nebul/models/moonshotai/Kimi-K3.toml
new file mode 100644
index 00000000000..91c5bd93f7e
--- /dev/null
+++ b/providers/nebul/models/moonshotai/Kimi-K3.toml
@@ -0,0 +1,14 @@
+base_model = "moonshotai/kimi-k3"
+
+# Controls per moonshot's official entry: thinking toggle + adaptive effort
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
+[cost]
+input = 4.73
+output = 23.63
+cache_read = 1.13
diff --git a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
new file mode 100644
index 00000000000..4ae1416245d
--- /dev/null
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
@@ -0,0 +1,10 @@
+base_model = "nvidia/nemotron-3-nano-30b-a3b"
+
+# Nemotron 3 hybrid reasoning: on/off toggle (deepinfra peer)
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 1
+output = 3
+cache_read = 0.5
diff --git a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
new file mode 100644
index 00000000000..9c849214092
--- /dev/null
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
@@ -0,0 +1,13 @@
+base_model = "nvidia/nemotron-3-super-120b-a12b"
+
+# Nemotron 3 hybrid reasoning: on/off toggle (crusoe peer)
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0.32
+output = 0.69
+cache_read = 0.08
+
+[limit]
+context = 1_000_000
diff --git a/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml b/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
new file mode 100644
index 00000000000..e987675f816
--- /dev/null
+++ b/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
@@ -0,0 +1,13 @@
+base_model = "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"
+
+# Nemotron 3 hybrid reasoning: on/off toggle (crusoe sibling pattern)
+[[reasoning_options]]
+type = "toggle"
+
+[cost]
+input = 0.16
+output = 1.05
+cache_read = 0.05
+
+[limit]
+context = 262_144
diff --git a/providers/nebul/models/openai/gpt-oss-120b.toml b/providers/nebul/models/openai/gpt-oss-120b.toml
new file mode 100644
index 00000000000..8477b4b5638
--- /dev/null
+++ b/providers/nebul/models/openai/gpt-oss-120b.toml
@@ -0,0 +1,11 @@
+base_model = "openai/gpt-oss-120b"
+
+# Efforts: reasoning_effort = low|medium|high — the model's documented Chat Completions control
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high"]
+
+[cost]
+input = 0.17
+output = 0.63
+cache_read = 0.05
diff --git a/providers/nebul/models/zai-org/GLM-5.3-Flash.toml b/providers/nebul/models/zai-org/GLM-5.3-Flash.toml
new file mode 100644
index 00000000000..ba6964468d5
--- /dev/null
+++ b/providers/nebul/models/zai-org/GLM-5.3-Flash.toml
@@ -0,0 +1,13 @@
+base_model = "zhipuai/glm-5.3-flash"
+
+# Efforts: reasoning_effort = low|high|max on POST /v1/chat/completions (verified live 2026-09-09)
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
+[cost]
+input = 0.21
+output = 0.63
+
+[limit]
+context = 1_048_572
diff --git a/providers/nebul/models/zai-org/GLM-5.3.toml b/providers/nebul/models/zai-org/GLM-5.3.toml
new file mode 100644
index 00000000000..b1f38252c8e
--- /dev/null
+++ b/providers/nebul/models/zai-org/GLM-5.3.toml
@@ -0,0 +1,14 @@
+base_model = "zhipuai/glm-5.3"
+
+# GLM-5.3 always reasons; efforts low|high|max per zhipuai's official entry
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
+[cost]
+input = 1.47
+output = 4.62
+cache_read = 0.35
+
+[limit]
+context = 1_048_576
diff --git a/providers/nebul/provider.toml b/providers/nebul/provider.toml
new file mode 100644
index 00000000000..dc44e31d87d
--- /dev/null
+++ b/providers/nebul/provider.toml
@@ -0,0 +1,5 @@
+name = "Nebul"
+npm = "@ai-sdk/openai-compatible"
+api = "https://api.inference.nebul.io/v1"
+env = ["NEBUL_API_KEY"]
+doc = "https://docs.nebul.io"

From 70b6735a90e4a2a1d5f6b8f7502818514fde8e25 Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 14:48:47 +0200
Subject: [PATCH 02/13] fix(mistral): mark Ministral 3 14B Instruct 2512 as
 open weights (Apache 2.0)

---
 models/mistral/ministral-3-14b-instruct-2512.toml | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/models/mistral/ministral-3-14b-instruct-2512.toml b/models/mistral/ministral-3-14b-instruct-2512.toml
index 7cd61ebb5a3..c027b18cab5 100644
--- a/models/mistral/ministral-3-14b-instruct-2512.toml
+++ b/models/mistral/ministral-3-14b-instruct-2512.toml
@@ -1,3 +1,7 @@
+# Sources (accessed 2026-09-11):
+# https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512 (Apache 2.0, public safetensors)
+# https://mistral.ai/news/mistral-3
+
 name = "Ministral 3 14B Instruct"
 description = "Efficient Mistral edge model balancing speed and cost for general-purpose tasks"
 family = "ministral"
@@ -8,7 +12,8 @@ reasoning = false
 temperature = true
 tool_call = true
 structured_output = true
-open_weights = false
+open_weights = true
+license = "Apache 2.0"
 
 [limit]
 context = 262_144
@@ -17,3 +22,7 @@ output = 262_144
 [modalities]
 input = ["text", "image"]
 output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512"

From 1239e4407200191c5dd959f2777c8b8ab5225065 Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 14:48:47 +0200
Subject: [PATCH 03/13] fix(nebul): drop reasoning controls not documented on
 Nebul's API

---
 providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml     |  9 +++------
 providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml       | 10 +++-------
 .../nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml  |  4 +++-
 providers/nebul/models/google/gemma-4-31B-it.toml      |  5 +++--
 .../nebul/models/meta-models/muse-glimmer-30b.toml     |  7 +++----
 providers/nebul/models/moonshotai/Kimi-K3.toml         | 10 +++-------
 .../nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml    |  6 +++---
 .../nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml |  6 +++---
 .../Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml    |  6 +++---
 9 files changed, 27 insertions(+), 36 deletions(-)

diff --git a/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml b/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
index b9cc203d0a9..1ff8b5e0bd1 100644
--- a/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
+++ b/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
@@ -1,11 +1,8 @@
+# Model reasons always-on; Nebul exposes no thinking toggle or budget field (its
+# docs cover reasoning_effort only): https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "alibaba/qwen3.5-397b-a17b"
 
-# Controls per alibaba's official entry: thinking toggle + budget
-[[reasoning_options]]
-type = "toggle"
-
-[[reasoning_options]]
-type = "budget_tokens"
+reasoning_options = []
 
 [cost]
 input = 0.63
diff --git a/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
index e0d778d4b2a..ed940b6af30 100644
--- a/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
+++ b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
@@ -1,12 +1,8 @@
+# Model reasons always-on; /model/info advertises no reasoning_efforts for it and
+# Nebul's docs cover reasoning_effort only: https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "alibaba/qwen3.8-27b"
 
-# Controls per deepinfra peer: thinking toggle + efforts
-[[reasoning_options]]
-type = "toggle"
-
-[[reasoning_options]]
-type = "effort"
-values = ["low", "medium", "xhigh"]
+reasoning_options = []
 
 [cost]
 input = 0.21
diff --git a/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml b/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
index d00053543cb..f9a3906bb2b 100644
--- a/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
+++ b/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
@@ -1,6 +1,8 @@
 base_model = "deepseek/deepseek-v4.1-flash"
 
-# Efforts: reasoning_effort = low|high|max (from /model/info reasoning_efforts)
+# Efforts: reasoning_effort = low|high|max (advertised by /model/info reasoning_efforts).
+# reasoning_effort is Nebul's only documented reasoning control; no separate on/off toggle.
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 [[reasoning_options]]
 type = "effort"
 values = ["low", "high", "max"]
diff --git a/providers/nebul/models/google/gemma-4-31B-it.toml b/providers/nebul/models/google/gemma-4-31B-it.toml
index 4c8e563df73..b02d4867f35 100644
--- a/providers/nebul/models/google/gemma-4-31B-it.toml
+++ b/providers/nebul/models/google/gemma-4-31B-it.toml
@@ -1,7 +1,8 @@
+# Model reasons always-on; Nebul exposes no caller control for it (its docs cover
+# reasoning_effort only): https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "google/gemma-4-31b-it"
 
-[[reasoning_options]]
-type = "toggle"
+reasoning_options = []
 
 [cost]
 input = 0.42
diff --git a/providers/nebul/models/meta-models/muse-glimmer-30b.toml b/providers/nebul/models/meta-models/muse-glimmer-30b.toml
index 0daac5a89cc..7f04f392040 100644
--- a/providers/nebul/models/meta-models/muse-glimmer-30b.toml
+++ b/providers/nebul/models/meta-models/muse-glimmer-30b.toml
@@ -1,9 +1,8 @@
+# Model reasons always-on; /model/info advertises no reasoning_efforts for it and
+# Nebul's docs cover reasoning_effort only: https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "meta/muse-glimmer-30b"
 
-# Efforts per fireworks peer (models without API-exposed levels keep lab controls)
-[[reasoning_options]]
-type = "effort"
-values = ["low", "medium", "high", "xhigh"]
+reasoning_options = []
 
 [cost]
 input = 0.7
diff --git a/providers/nebul/models/moonshotai/Kimi-K3.toml b/providers/nebul/models/moonshotai/Kimi-K3.toml
index 91c5bd93f7e..b9e9431a442 100644
--- a/providers/nebul/models/moonshotai/Kimi-K3.toml
+++ b/providers/nebul/models/moonshotai/Kimi-K3.toml
@@ -1,12 +1,8 @@
+# Model reasons always-on; /model/info advertises no reasoning_efforts for it and
+# Nebul's docs cover reasoning_effort only: https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "moonshotai/kimi-k3"
 
-# Controls per moonshot's official entry: thinking toggle + adaptive effort
-[[reasoning_options]]
-type = "toggle"
-
-[[reasoning_options]]
-type = "effort"
-values = ["low", "high", "max"]
+reasoning_options = []
 
 [cost]
 input = 4.73
diff --git a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
index 4ae1416245d..b178db29362 100644
--- a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
@@ -1,8 +1,8 @@
+# Hybrid reasoning; Nebul exposes no caller control for it (its docs cover
+# reasoning_effort only): https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "nvidia/nemotron-3-nano-30b-a3b"
 
-# Nemotron 3 hybrid reasoning: on/off toggle (deepinfra peer)
-[[reasoning_options]]
-type = "toggle"
+reasoning_options = []
 
 [cost]
 input = 1
diff --git a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
index 9c849214092..b83ee9731f3 100644
--- a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
@@ -1,8 +1,8 @@
+# Hybrid reasoning; Nebul exposes no caller control for it (its docs cover
+# reasoning_effort only): https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "nvidia/nemotron-3-super-120b-a12b"
 
-# Nemotron 3 hybrid reasoning: on/off toggle (crusoe peer)
-[[reasoning_options]]
-type = "toggle"
+reasoning_options = []
 
 [cost]
 input = 0.32
diff --git a/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml b/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
index e987675f816..c2547361c3a 100644
--- a/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
+++ b/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
@@ -1,8 +1,8 @@
+# Hybrid reasoning; Nebul exposes no caller control for it (its docs cover
+# reasoning_effort only): https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"
 
-# Nemotron 3 hybrid reasoning: on/off toggle (crusoe sibling pattern)
-[[reasoning_options]]
-type = "toggle"
+reasoning_options = []
 
 [cost]
 input = 0.16

From 4717cca8d9b35bf62a1c4e0e941f2bc7f4aed65d Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 14:48:47 +0200
Subject: [PATCH 04/13] fix(nebul): preserve hand-authored non-effort reasoning
 controls during sync

---
 packages/core/src/sync/providers/nebul.ts |  5 ++++-
 packages/core/test/nebul.test.ts          | 14 ++++++++++++++
 2 files changed, 18 insertions(+), 1 deletion(-)

diff --git a/packages/core/src/sync/providers/nebul.ts b/packages/core/src/sync/providers/nebul.ts
index ea29002e259..883c6f5ef91 100644
--- a/packages/core/src/sync/providers/nebul.ts
+++ b/packages/core/src/sync/providers/nebul.ts
@@ -117,10 +117,13 @@ function isCatalogChatModel(entry: NebulEntry): boolean {
     && entry.model_name !== PING_MODEL && !DEPRECATED.has(entry.model_name) && !DENYLIST.test(entry.model_name);
 }
 
+// /model/info only describes the effort control; hand-authored toggle and
+// budget_tokens options must survive rewrites, so replace just the effort entry.
 function buildReasoningOptions(entry: NebulEntry, existing: ExistingModel | undefined) {
   const efforts = entry.model_info.reasoning_efforts ?? [];
   if (efforts.length === 0) return existing?.reasoning_options;
-  return [{ type: "effort" as const, values: efforts }];
+  const preserved = existing?.reasoning_options?.filter((option) => option.type !== "effort") ?? [];
+  return [...preserved, { type: "effort" as const, values: efforts }];
 }
 
 function resolveBaseModel(servedID: string, huggingfaceID: string | undefined): string | undefined {
diff --git a/packages/core/test/nebul.test.ts b/packages/core/test/nebul.test.ts
index e2d2aed1fc1..88caba0a9d3 100644
--- a/packages/core/test/nebul.test.ts
+++ b/packages/core/test/nebul.test.ts
@@ -50,6 +50,20 @@ test("preserves authored reasoning controls when the host exposes no efforts", (
   expect(translated?.model.reasoning_options).toEqual(authored);
 });
 
+test("preserves authored toggle and budget controls when the host advertises efforts", () => {
+  const authored = [
+    { type: "toggle" as const },
+    { type: "budget_tokens" as const },
+    { type: "effort" as const, values: ["low"] },
+  ];
+  const translated = nebul.translateModel(nebulEntry("zai-org/GLM-5.3"), context(existingWith(authored)));
+  expect(translated?.model.reasoning_options).toEqual([
+    { type: "toggle" },
+    { type: "budget_tokens" },
+    { type: "effort", values: ["low", "high", "max"] },
+  ]);
+});
+
 test("resolves base models across org renames and quantization suffixes", () => {
   const cases: [string, string | null, string][] = [
     ["Qwen/Qwen3.8-27B-FP8", "Qwen/Qwen3.8-27B-FP8", "alibaba/qwen3.8-27b"],

From 09a48a9e9a3b9bc066c7c9d2d3b6a01f6f5af7fe Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 15:02:52 +0200
Subject: [PATCH 05/13] fix(nebul): keep hand-authored models when source data
 is incomplete

---
 packages/core/src/sync/providers/nebul.ts | 41 +++++++++++++++--------
 packages/core/test/nebul.test.ts          | 26 ++++++++++++++
 2 files changed, 53 insertions(+), 14 deletions(-)

diff --git a/packages/core/src/sync/providers/nebul.ts b/packages/core/src/sync/providers/nebul.ts
index 883c6f5ef91..a5b5bf408fc 100644
--- a/packages/core/src/sync/providers/nebul.ts
+++ b/packages/core/src/sync/providers/nebul.ts
@@ -80,22 +80,35 @@ export const nebul = {
     if (!isCatalogChatModel(entry)) return undefined;
     const id = entry.model_name;
     const info = entry.model_info;
-    const baseModel = resolveBaseModel(id, info.huggingface_id ?? undefined);
-    if (baseModel === undefined) return undefined;
-    if (info.input_cost_per_1m_tokens == null || info.output_cost_per_1m_tokens == null || info.max_input_tokens == null) return undefined;
+    const existing = context.existing(id);
+    // Existing entries must survive incomplete source data — a transient null
+    // price or an unresolved alias would otherwise delete the hand-authored
+    // TOML on the next run. They keep their authored base_model and cost/limit;
+    // only brand-new models need a fully-priced, resolvable source entry.
+    const baseModel = existing?.base_model ?? resolveBaseModel(id, info.huggingface_id ?? undefined);
+    const cost = info.input_cost_per_1m_tokens != null && info.output_cost_per_1m_tokens != null
+      ? {
+          input: info.input_cost_per_1m_tokens,
+          output: info.output_cost_per_1m_tokens,
+          cache_read: info.cache_read_input_cost_per_1m_tokens ?? undefined,
+        }
+      : existing?.cost;
+    const limit = info.max_input_tokens != null ? { context: info.max_input_tokens } : existing?.limit;
+    if (existing === undefined && (baseModel === undefined || cost === undefined || limit === undefined)) return undefined;
     const values = {
-      reasoning_options: buildReasoningOptions(entry, context.existing(id)),
-      cost: {
-        input: info.input_cost_per_1m_tokens,
-        output: info.output_cost_per_1m_tokens,
-        cache_read: info.cache_read_input_cost_per_1m_tokens ?? undefined,
-      },
-      limit: { context: info.max_input_tokens },
-    };
-    return {
-      id,
-      model: factorBaseModel(baseModel, values, values.limit) as SyncedModel,
+      reasoning_options: buildReasoningOptions(entry, existing),
+      cost,
+      limit,
     };
+    if (baseModel !== undefined) {
+      return {
+        id,
+        model: factorBaseModel(baseModel, values, limit) as SyncedModel,
+      };
+    }
+    // Existing standalone definition whose served alias no longer resolves:
+    // keep the authored fields, refreshing only what /model/info still provides.
+    return { id, model: { ...existing, ...values } as SyncedModel };
   },
   // Only report chat models whose base_model could not be resolved; filtered
   // serving artifacts (embeddings, rerankers, the ping model) skip silently.
diff --git a/packages/core/test/nebul.test.ts b/packages/core/test/nebul.test.ts
index 88caba0a9d3..43b020113d6 100644
--- a/packages/core/test/nebul.test.ts
+++ b/packages/core/test/nebul.test.ts
@@ -64,6 +64,32 @@ test("preserves authored toggle and budget controls when the host advertises eff
   ]);
 });
 
+test("keeps existing entries when the source pricing or context is temporarily null", () => {
+  const existing = {
+    base_model: "zhipuai/glm-5.3",
+    cost: { input: 1.47, output: 4.62 },
+    limit: { context: 1_048_576 },
+  } as ExistingModel;
+  const translated = nebul.translateModel(
+    nebulEntry("zai-org/GLM-5.3", { input_cost_per_1m_tokens: null, output_cost_per_1m_tokens: null, max_input_tokens: null }),
+    context(existing),
+  );
+  expect(translated).toMatchObject({
+    id: "zai-org/GLM-5.3",
+    model: { base_model: "zhipuai/glm-5.3", cost: { input: 1.47, output: 4.62 }, limit: { context: 1_048_576 } },
+  });
+});
+
+test("keeps existing entries when the served alias no longer resolves to lab metadata", () => {
+  const existing = {
+    base_model: "zhipuai/glm-5.3",
+    cost: { input: 1.47, output: 4.62 },
+    limit: { context: 1_048_576 },
+  } as ExistingModel;
+  const translated = nebul.translateModel(nebulEntry("someorg/Unknown-Model", { huggingface_id: null }), context(existing));
+  expect(translated?.model.base_model).toBe("zhipuai/glm-5.3");
+});
+
 test("resolves base models across org renames and quantization suffixes", () => {
   const cases: [string, string | null, string][] = [
     ["Qwen/Qwen3.8-27B-FP8", "Qwen/Qwen3.8-27B-FP8", "alibaba/qwen3.8-27b"],

From ad090c8662740942c38c9ab629a7dd0cb45f2b2f Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 15:02:53 +0200
Subject: [PATCH 06/13] fix(nebul): cite /model/info as the source of
 Mistral-Medium-3.5 effort values

---
 providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml b/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
index 331f2d4f281..e00a2b8a97a 100644
--- a/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
+++ b/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
@@ -1,6 +1,8 @@
 base_model = "mistral/mistral-medium-2604"
 
-# Efforts: reasoning_effort = low|medium|high|max (from /model/info reasoning_efforts)
+# Efforts: reasoning_effort = low|medium|high|max — exactly /model/info's advertised
+# reasoning_efforts for this model, superseding the lab's none|high set.
+# https://api.inference.nebul.io/model/info
 [[reasoning_options]]
 type = "effort"
 values = ["low", "medium", "high", "max"]

From 6e139d66741fe07e26ef342fd281364dc04d19a2 Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 15:02:53 +0200
Subject: [PATCH 07/13] fix(mistral): cite HF card for the Ministral output
 limit

---
 models/mistral/ministral-3-14b-instruct-2512.toml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/models/mistral/ministral-3-14b-instruct-2512.toml b/models/mistral/ministral-3-14b-instruct-2512.toml
index c027b18cab5..335a18814f2 100644
--- a/models/mistral/ministral-3-14b-instruct-2512.toml
+++ b/models/mistral/ministral-3-14b-instruct-2512.toml
@@ -1,6 +1,7 @@
 # Sources (accessed 2026-09-11):
 # https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512 (Apache 2.0, public safetensors)
 # https://mistral.ai/news/mistral-3
+# limit.output 262_144 per the HF card's own usage example (MAX_TOK = max_tokens = 262144).
 
 name = "Ministral 3 14B Instruct"
 description = "Efficient Mistral edge model balancing speed and cost for general-purpose tasks"

From 385e7bea138dfc975ebe868ce11259682218656e Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 15:45:28 +0200
Subject: [PATCH 08/13] fix(nebul): keep sync-owned source comments in leading
 header blocks

---
 providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml   | 4 ++--
 providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml | 4 ++--
 providers/nebul/models/openai/gpt-oss-120b.toml               | 2 +-
 providers/nebul/models/zai-org/GLM-5.3-Flash.toml             | 2 +-
 providers/nebul/models/zai-org/GLM-5.3.toml                   | 2 +-
 5 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml b/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
index f9a3906bb2b..b1e92a13055 100644
--- a/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
+++ b/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
@@ -1,8 +1,8 @@
-base_model = "deepseek/deepseek-v4.1-flash"
-
 # Efforts: reasoning_effort = low|high|max (advertised by /model/info reasoning_efforts).
 # reasoning_effort is Nebul's only documented reasoning control; no separate on/off toggle.
 # https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+base_model = "deepseek/deepseek-v4.1-flash"
+
 [[reasoning_options]]
 type = "effort"
 values = ["low", "high", "max"]
diff --git a/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml b/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
index e00a2b8a97a..c790f7c2f37 100644
--- a/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
+++ b/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
@@ -1,8 +1,8 @@
-base_model = "mistral/mistral-medium-2604"
-
 # Efforts: reasoning_effort = low|medium|high|max — exactly /model/info's advertised
 # reasoning_efforts for this model, superseding the lab's none|high set.
 # https://api.inference.nebul.io/model/info
+base_model = "mistral/mistral-medium-2604"
+
 [[reasoning_options]]
 type = "effort"
 values = ["low", "medium", "high", "max"]
diff --git a/providers/nebul/models/openai/gpt-oss-120b.toml b/providers/nebul/models/openai/gpt-oss-120b.toml
index 8477b4b5638..e3aab6851f9 100644
--- a/providers/nebul/models/openai/gpt-oss-120b.toml
+++ b/providers/nebul/models/openai/gpt-oss-120b.toml
@@ -1,6 +1,6 @@
+# Efforts: reasoning_effort = low|medium|high — the model's documented Chat Completions control
 base_model = "openai/gpt-oss-120b"
 
-# Efforts: reasoning_effort = low|medium|high — the model's documented Chat Completions control
 [[reasoning_options]]
 type = "effort"
 values = ["low", "medium", "high"]
diff --git a/providers/nebul/models/zai-org/GLM-5.3-Flash.toml b/providers/nebul/models/zai-org/GLM-5.3-Flash.toml
index ba6964468d5..eef2f917969 100644
--- a/providers/nebul/models/zai-org/GLM-5.3-Flash.toml
+++ b/providers/nebul/models/zai-org/GLM-5.3-Flash.toml
@@ -1,6 +1,6 @@
+# Efforts: reasoning_effort = low|high|max on POST /v1/chat/completions (verified live 2026-09-09)
 base_model = "zhipuai/glm-5.3-flash"
 
-# Efforts: reasoning_effort = low|high|max on POST /v1/chat/completions (verified live 2026-09-09)
 [[reasoning_options]]
 type = "effort"
 values = ["low", "high", "max"]
diff --git a/providers/nebul/models/zai-org/GLM-5.3.toml b/providers/nebul/models/zai-org/GLM-5.3.toml
index b1f38252c8e..937d438b7fe 100644
--- a/providers/nebul/models/zai-org/GLM-5.3.toml
+++ b/providers/nebul/models/zai-org/GLM-5.3.toml
@@ -1,6 +1,6 @@
+# GLM-5.3 always reasons; efforts low|high|max per zhipuai's official entry
 base_model = "zhipuai/glm-5.3"
 
-# GLM-5.3 always reasons; efforts low|high|max per zhipuai's official entry
 [[reasoning_options]]
 type = "effort"
 values = ["low", "high", "max"]

From f3a1584fc642455c9da3bccfadc41c74270bb696 Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 16:03:19 +0200
Subject: [PATCH 09/13] fix(nebul): author interleaved reasoning side channels

Nebul's docs map each family to a trace channel: reasoning_content for
OpenAI-style models, the Nemotron family, Qwen3 thinking variants (plus
DeepSeek-V4.1-Flash, Kimi-K3, Mistral-Medium-3.5 and muse-glimmer-30b,
confirmed against the serving stack), message.reasoning for GLM-5.x and
inline <thinking> for Gemma. Author interleaved on all 13 reasoners
(interleaved = true where the schema has no named field) and carry
existing?.interleaved through translateModel so sync never drops it.
---
 packages/core/src/sync/providers/nebul.ts             |  1 +
 packages/core/test/nebul.test.ts                      | 11 +++++++++++
 providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml    |  3 +++
 providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml      |  3 +++
 .../nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml |  3 +++
 providers/nebul/models/google/gemma-4-31B-it.toml     |  3 +++
 .../nebul/models/meta-models/muse-glimmer-30b.toml    |  3 +++
 .../models/mistralai/Mistral-Medium-3.5-128B.toml     |  3 +++
 providers/nebul/models/moonshotai/Kimi-K3.toml        |  3 +++
 .../nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml   |  3 +++
 .../NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml       |  3 +++
 .../Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml   |  3 +++
 providers/nebul/models/openai/gpt-oss-120b.toml       |  3 +++
 providers/nebul/models/zai-org/GLM-5.3-Flash.toml     |  3 +++
 providers/nebul/models/zai-org/GLM-5.3.toml           |  3 +++
 15 files changed, 51 insertions(+)

diff --git a/packages/core/src/sync/providers/nebul.ts b/packages/core/src/sync/providers/nebul.ts
index a5b5bf408fc..6088cce1b9b 100644
--- a/packages/core/src/sync/providers/nebul.ts
+++ b/packages/core/src/sync/providers/nebul.ts
@@ -96,6 +96,7 @@ export const nebul = {
     const limit = info.max_input_tokens != null ? { context: info.max_input_tokens } : existing?.limit;
     if (existing === undefined && (baseModel === undefined || cost === undefined || limit === undefined)) return undefined;
     const values = {
+      interleaved: existing?.interleaved,
       reasoning_options: buildReasoningOptions(entry, existing),
       cost,
       limit,
diff --git a/packages/core/test/nebul.test.ts b/packages/core/test/nebul.test.ts
index 43b020113d6..543fd76fa87 100644
--- a/packages/core/test/nebul.test.ts
+++ b/packages/core/test/nebul.test.ts
@@ -64,6 +64,17 @@ test("preserves authored toggle and budget controls when the host advertises eff
   ]);
 });
 
+test("carries authored interleaved through sync", () => {
+  const inline = nebul.translateModel(nebulEntry("zai-org/GLM-5.3"), context({ interleaved: true } as ExistingModel));
+  expect(inline?.model.interleaved).toBe(true);
+
+  const named = nebul.translateModel(
+    nebulEntry("deepseek-ai/DeepSeek-V4.1-Flash"),
+    context({ interleaved: { field: "reasoning_content" } } as ExistingModel),
+  );
+  expect(named?.model.interleaved).toEqual({ field: "reasoning_content" });
+});
+
 test("keeps existing entries when the source pricing or context is temporarily null", () => {
   const existing = {
     base_model: "zhipuai/glm-5.3",
diff --git a/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml b/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
index 1ff8b5e0bd1..19509361dbc 100644
--- a/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
+++ b/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
@@ -4,6 +4,9 @@ base_model = "alibaba/qwen3.5-397b-a17b"
 
 reasoning_options = []
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.63
 output = 3.78
diff --git a/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
index ed940b6af30..b5588318994 100644
--- a/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
+++ b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
@@ -4,6 +4,9 @@ base_model = "alibaba/qwen3.8-27b"
 
 reasoning_options = []
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.21
 output = 0.74
diff --git a/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml b/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
index b1e92a13055..c39e6cb1602 100644
--- a/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
+++ b/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
@@ -3,6 +3,9 @@
 # https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "deepseek/deepseek-v4.1-flash"
 
+[interleaved]
+field = "reasoning_content"
+
 [[reasoning_options]]
 type = "effort"
 values = ["low", "high", "max"]
diff --git a/providers/nebul/models/google/gemma-4-31B-it.toml b/providers/nebul/models/google/gemma-4-31B-it.toml
index b02d4867f35..756e52c6fa5 100644
--- a/providers/nebul/models/google/gemma-4-31B-it.toml
+++ b/providers/nebul/models/google/gemma-4-31B-it.toml
@@ -1,9 +1,12 @@
 # Model reasons always-on; Nebul exposes no caller control for it (its docs cover
 # reasoning_effort only): https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+# Traces arrive inline as <thinking> inside the content (Nebul docs), so interleaved = true
 base_model = "google/gemma-4-31b-it"
 
 reasoning_options = []
 
+interleaved = true
+
 [cost]
 input = 0.42
 output = 1.05
diff --git a/providers/nebul/models/meta-models/muse-glimmer-30b.toml b/providers/nebul/models/meta-models/muse-glimmer-30b.toml
index 7f04f392040..a224048aadf 100644
--- a/providers/nebul/models/meta-models/muse-glimmer-30b.toml
+++ b/providers/nebul/models/meta-models/muse-glimmer-30b.toml
@@ -4,6 +4,9 @@ base_model = "meta/muse-glimmer-30b"
 
 reasoning_options = []
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.7
 output = 3.5
diff --git a/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml b/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
index c790f7c2f37..428929c4918 100644
--- a/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
+++ b/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
@@ -3,6 +3,9 @@
 # https://api.inference.nebul.io/model/info
 base_model = "mistral/mistral-medium-2604"
 
+[interleaved]
+field = "reasoning_content"
+
 [[reasoning_options]]
 type = "effort"
 values = ["low", "medium", "high", "max"]
diff --git a/providers/nebul/models/moonshotai/Kimi-K3.toml b/providers/nebul/models/moonshotai/Kimi-K3.toml
index b9e9431a442..2546ee5932f 100644
--- a/providers/nebul/models/moonshotai/Kimi-K3.toml
+++ b/providers/nebul/models/moonshotai/Kimi-K3.toml
@@ -4,6 +4,9 @@ base_model = "moonshotai/kimi-k3"
 
 reasoning_options = []
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 4.73
 output = 23.63
diff --git a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
index b178db29362..c3018447d6d 100644
--- a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
@@ -4,6 +4,9 @@ base_model = "nvidia/nemotron-3-nano-30b-a3b"
 
 reasoning_options = []
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 1
 output = 3
diff --git a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
index b83ee9731f3..abead25f50b 100644
--- a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
@@ -4,6 +4,9 @@ base_model = "nvidia/nemotron-3-super-120b-a12b"
 
 reasoning_options = []
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.32
 output = 0.69
diff --git a/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml b/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
index c2547361c3a..94094ee7c95 100644
--- a/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
+++ b/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
@@ -4,6 +4,9 @@ base_model = "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"
 
 reasoning_options = []
 
+[interleaved]
+field = "reasoning_content"
+
 [cost]
 input = 0.16
 output = 1.05
diff --git a/providers/nebul/models/openai/gpt-oss-120b.toml b/providers/nebul/models/openai/gpt-oss-120b.toml
index e3aab6851f9..7e53cb3b259 100644
--- a/providers/nebul/models/openai/gpt-oss-120b.toml
+++ b/providers/nebul/models/openai/gpt-oss-120b.toml
@@ -1,6 +1,9 @@
 # Efforts: reasoning_effort = low|medium|high — the model's documented Chat Completions control
 base_model = "openai/gpt-oss-120b"
 
+[interleaved]
+field = "reasoning_content"
+
 [[reasoning_options]]
 type = "effort"
 values = ["low", "medium", "high"]
diff --git a/providers/nebul/models/zai-org/GLM-5.3-Flash.toml b/providers/nebul/models/zai-org/GLM-5.3-Flash.toml
index eef2f917969..892df7195e5 100644
--- a/providers/nebul/models/zai-org/GLM-5.3-Flash.toml
+++ b/providers/nebul/models/zai-org/GLM-5.3-Flash.toml
@@ -1,6 +1,9 @@
 # Efforts: reasoning_effort = low|high|max on POST /v1/chat/completions (verified live 2026-09-09)
+# Traces arrive in message.reasoning (Nebul docs), so interleaved = true
 base_model = "zhipuai/glm-5.3-flash"
 
+interleaved = true
+
 [[reasoning_options]]
 type = "effort"
 values = ["low", "high", "max"]
diff --git a/providers/nebul/models/zai-org/GLM-5.3.toml b/providers/nebul/models/zai-org/GLM-5.3.toml
index 937d438b7fe..0a964c0dda7 100644
--- a/providers/nebul/models/zai-org/GLM-5.3.toml
+++ b/providers/nebul/models/zai-org/GLM-5.3.toml
@@ -1,6 +1,9 @@
 # GLM-5.3 always reasons; efforts low|high|max per zhipuai's official entry
+# Traces arrive in message.reasoning (Nebul docs), so interleaved = true
 base_model = "zhipuai/glm-5.3"
 
+interleaved = true
+
 [[reasoning_options]]
 type = "effort"
 values = ["low", "high", "max"]

From ed484bf2fa5b430654bb320953687129a184c3dc Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 16:03:26 +0200
Subject: [PATCH 10/13] fix(nebul): sync writes effort-only options when the
 host advertises efforts
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Nebul's API has no toggle or budget control, so lab-style options must
not be merged back into a rewritten entry — that would claim wire fields
Nebul does not accept. When reasoning_efforts is advertised, the effort
entry is the whole set; when it is not, authored options are kept.
---
 packages/core/src/sync/providers/nebul.ts | 9 +++++----
 packages/core/test/nebul.test.ts          | 8 ++------
 2 files changed, 7 insertions(+), 10 deletions(-)

diff --git a/packages/core/src/sync/providers/nebul.ts b/packages/core/src/sync/providers/nebul.ts
index 6088cce1b9b..3470534d554 100644
--- a/packages/core/src/sync/providers/nebul.ts
+++ b/packages/core/src/sync/providers/nebul.ts
@@ -131,13 +131,14 @@ function isCatalogChatModel(entry: NebulEntry): boolean {
     && entry.model_name !== PING_MODEL && !DEPRECATED.has(entry.model_name) && !DENYLIST.test(entry.model_name);
 }
 
-// /model/info only describes the effort control; hand-authored toggle and
-// budget_tokens options must survive rewrites, so replace just the effort entry.
+// Nebul documents exactly one reasoning control: reasoning_effort. When the
+// host advertises efforts, write that effort entry and nothing else —
+// lab-style toggles or budgets are not supported on this API. When it
+// advertises none, keep the authored options.
 function buildReasoningOptions(entry: NebulEntry, existing: ExistingModel | undefined) {
   const efforts = entry.model_info.reasoning_efforts ?? [];
   if (efforts.length === 0) return existing?.reasoning_options;
-  const preserved = existing?.reasoning_options?.filter((option) => option.type !== "effort") ?? [];
-  return [...preserved, { type: "effort" as const, values: efforts }];
+  return [{ type: "effort" as const, values: efforts }];
 }
 
 function resolveBaseModel(servedID: string, huggingfaceID: string | undefined): string | undefined {
diff --git a/packages/core/test/nebul.test.ts b/packages/core/test/nebul.test.ts
index 543fd76fa87..ee84ec3165d 100644
--- a/packages/core/test/nebul.test.ts
+++ b/packages/core/test/nebul.test.ts
@@ -50,18 +50,14 @@ test("preserves authored reasoning controls when the host exposes no efforts", (
   expect(translated?.model.reasoning_options).toEqual(authored);
 });
 
-test("preserves authored toggle and budget controls when the host advertises efforts", () => {
+test("replaces authored options with the advertised effort entry when efforts are advertised", () => {
   const authored = [
     { type: "toggle" as const },
     { type: "budget_tokens" as const },
     { type: "effort" as const, values: ["low"] },
   ];
   const translated = nebul.translateModel(nebulEntry("zai-org/GLM-5.3"), context(existingWith(authored)));
-  expect(translated?.model.reasoning_options).toEqual([
-    { type: "toggle" },
-    { type: "budget_tokens" },
-    { type: "effort", values: ["low", "high", "max"] },
-  ]);
+  expect(translated?.model.reasoning_options).toEqual([{ type: "effort", values: ["low", "high", "max"] }]);
 });
 
 test("carries authored interleaved through sync", () => {

From 3b0171319c12bced2d21c3b8ce3d4edc555f2df1 Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 17:00:59 +0200
Subject: [PATCH 11/13] fix(nebul): author peer-derived effort sets for three
 unadvertised reasoners

Nebul's docs state reasoning_effort is accepted for reasoning-capable
models and ignored only by non-reasoning models, so an empty option set
would falsely claim no caller control. Author effort options from
same-surface peers intersected with Nebul's documented value set:
Kimi-K3 low|high|max (Tinfoil), muse-glimmer-30b low|medium|high
(Fireworks/Vercel, xhigh undocumented on Nebul), Qwen3.8-27B low|medium
(OpenRouter, toggle absent on Nebul).
---
 packages/core/test/nebul.test.ts                     |  9 +++++++++
 providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml     | 12 ++++++++----
 .../nebul/models/meta-models/muse-glimmer-30b.toml   | 12 ++++++++----
 providers/nebul/models/moonshotai/Kimi-K3.toml       | 12 ++++++++----
 4 files changed, 33 insertions(+), 12 deletions(-)

diff --git a/packages/core/test/nebul.test.ts b/packages/core/test/nebul.test.ts
index ee84ec3165d..77f76e53f3c 100644
--- a/packages/core/test/nebul.test.ts
+++ b/packages/core/test/nebul.test.ts
@@ -71,6 +71,15 @@ test("carries authored interleaved through sync", () => {
   expect(named?.model.interleaved).toEqual({ field: "reasoning_content" });
 });
 
+test("keeps authored effort sets when the host advertises none", () => {
+  const authored = [{ type: "effort" as const, values: ["low", "high", "max"] }];
+  const translated = nebul.translateModel(
+    nebulEntry("moonshotai/Kimi-K3", { reasoning_efforts: [] }),
+    context(existingWith(authored)),
+  );
+  expect(translated?.model.reasoning_options).toEqual(authored);
+});
+
 test("keeps existing entries when the source pricing or context is temporarily null", () => {
   const existing = {
     base_model: "zhipuai/glm-5.3",
diff --git a/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
index b5588318994..c86aafb7ef4 100644
--- a/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
+++ b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
@@ -1,12 +1,16 @@
-# Model reasons always-on; /model/info advertises no reasoning_efforts for it and
-# Nebul's docs cover reasoning_effort only: https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+# Effort: reasoning_effort (Nebul docs: accepted for reasoning-capable models, ignored
+# only by non-reasoning models). Set is OpenRouter's low|medium|xhigh intersected with
+# Nebul's documented value set; Nebul has no toggle and xhigh is undocumented on it.
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "alibaba/qwen3.8-27b"
 
-reasoning_options = []
-
 [interleaved]
 field = "reasoning_content"
 
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium"]
+
 [cost]
 input = 0.21
 output = 0.74
diff --git a/providers/nebul/models/meta-models/muse-glimmer-30b.toml b/providers/nebul/models/meta-models/muse-glimmer-30b.toml
index a224048aadf..b0b8eb26284 100644
--- a/providers/nebul/models/meta-models/muse-glimmer-30b.toml
+++ b/providers/nebul/models/meta-models/muse-glimmer-30b.toml
@@ -1,12 +1,16 @@
-# Model reasons always-on; /model/info advertises no reasoning_efforts for it and
-# Nebul's docs cover reasoning_effort only: https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+# Effort: reasoning_effort (Nebul docs: accepted for reasoning-capable models, ignored
+# only by non-reasoning models). Set is Fireworks/Vercel's low|medium|high|xhigh
+# intersected with Nebul's documented value set; xhigh is undocumented on Nebul.
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "meta/muse-glimmer-30b"
 
-reasoning_options = []
-
 [interleaved]
 field = "reasoning_content"
 
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high"]
+
 [cost]
 input = 0.7
 output = 3.5
diff --git a/providers/nebul/models/moonshotai/Kimi-K3.toml b/providers/nebul/models/moonshotai/Kimi-K3.toml
index 2546ee5932f..fe1d9661f50 100644
--- a/providers/nebul/models/moonshotai/Kimi-K3.toml
+++ b/providers/nebul/models/moonshotai/Kimi-K3.toml
@@ -1,12 +1,16 @@
-# Model reasons always-on; /model/info advertises no reasoning_efforts for it and
-# Nebul's docs cover reasoning_effort only: https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+# Effort: reasoning_effort (Nebul docs: accepted for reasoning-capable models, ignored
+# only by non-reasoning models). Set mirrors same-surface peer Tinfoil (low|high|max);
+# /model/info does not enumerate efforts for this model.
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "moonshotai/kimi-k3"
 
-reasoning_options = []
-
 [interleaved]
 field = "reasoning_content"
 
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
+
 [cost]
 input = 4.73
 output = 23.63

From d3bdd8086642f1fb70d574bb68473eccc1ec50ea Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 19:53:35 +0200
Subject: [PATCH 12/13] fix(nebul): author documented effort sets and fail
 closed on unadvertised reasoners

---
 .../ministral-3-14b-instruct-2512.toml        |  6 +++++-
 packages/core/src/sync/providers/nebul.ts     | 19 +++++++++++++++++--
 packages/core/test/nebul.test.ts              |  9 +++++++++
 .../nebul/models/Qwen/Qwen3.5-397B-A17B.toml  | 12 ++++++++----
 .../nebul/models/Qwen/Qwen3.8-27B-FP8.toml    |  6 +++---
 .../nebul/models/google/gemma-4-31B-it.toml   | 12 ++++++++----
 .../NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml  | 12 ++++++++----
 ...VIDIA-Nemotron-3-Super-120B-A12B-BF16.toml | 12 ++++++++----
 ...ron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml | 12 ++++++++----
 9 files changed, 74 insertions(+), 26 deletions(-)

diff --git a/models/mistral/ministral-3-14b-instruct-2512.toml b/models/mistral/ministral-3-14b-instruct-2512.toml
index 335a18814f2..946a426285c 100644
--- a/models/mistral/ministral-3-14b-instruct-2512.toml
+++ b/models/mistral/ministral-3-14b-instruct-2512.toml
@@ -1,7 +1,11 @@
 # Sources (accessed 2026-09-11):
 # https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512 (Apache 2.0, public safetensors)
 # https://mistral.ai/news/mistral-3
-# limit.output 262_144 per the HF card's own usage example (MAX_TOK = max_tokens = 262144).
+# limit.output 262_144: the HF card states a 256k (= 262,144) context window, its vLLM
+# section defaults --max-model-len to 262144, and its own examples request
+# max_tokens = 262144; Mistral publishes no separate generation cap. Current-gen Mistral
+# flagships in this catalog (mistral-large-2512, mistral-medium-2604, devstral-2512) set
+# the same output = 262_144.
 
 name = "Ministral 3 14B Instruct"
 description = "Efficient Mistral edge model balancing speed and cost for general-purpose tasks"
diff --git a/packages/core/src/sync/providers/nebul.ts b/packages/core/src/sync/providers/nebul.ts
index 3470534d554..51bac2061dc 100644
--- a/packages/core/src/sync/providers/nebul.ts
+++ b/packages/core/src/sync/providers/nebul.ts
@@ -3,7 +3,8 @@ import path from "node:path";
 import { z } from "zod";
 
 import type { ExistingModel, SyncProvider, SyncedModel } from "../index.js";
-import { factorBaseModel } from "./openrouter.js";
+import { MissingReasoningOptionsError } from "../missing-reasoning-options.js";
+import { factorBaseModel, modelMetadata } from "./openrouter.js";
 
 const API_ENDPOINT = "https://api.inference.nebul.io/model/info";
 const MODELS_DIR = path.join(import.meta.dirname, "..", "..", "..", "..", "..", "models");
@@ -95,6 +96,19 @@ export const nebul = {
       : existing?.cost;
     const limit = info.max_input_tokens != null ? { context: info.max_input_tokens } : existing?.limit;
     if (existing === undefined && (baseModel === undefined || cost === undefined || limit === undefined)) return undefined;
+    // Fail closed rather than emitting no reasoning_options: a reasoner with
+    // neither advertised efforts nor authored options would sync as an empty
+    // entry (no caller control). The runner keeps the file and lists it in the
+    // skipped notice so the options can be hand-authored.
+    const isReasoner = baseModel !== undefined
+      ? modelMetadata(baseModel).reasoning === true
+      : existing?.reasoning === true;
+    if (isReasoner && (info.reasoning_efforts ?? []).length === 0 && existing?.reasoning_options === undefined) {
+      throw new MissingReasoningOptionsError(
+        id,
+        `${id} is a reasoning model, but Nebul advertises no reasoning_efforts and the catalog entry has no reasoning_options; hand-author them`,
+      );
+    }
     const values = {
       interleaved: existing?.interleaved,
       reasoning_options: buildReasoningOptions(entry, existing),
@@ -134,7 +148,8 @@ function isCatalogChatModel(entry: NebulEntry): boolean {
 // Nebul documents exactly one reasoning control: reasoning_effort. When the
 // host advertises efforts, write that effort entry and nothing else —
 // lab-style toggles or budgets are not supported on this API. When it
-// advertises none, keep the authored options.
+// advertises none, keep the authored options; a reasoner with neither is
+// rejected above so no empty options entry is ever synced.
 function buildReasoningOptions(entry: NebulEntry, existing: ExistingModel | undefined) {
   const efforts = entry.model_info.reasoning_efforts ?? [];
   if (efforts.length === 0) return existing?.reasoning_options;
diff --git a/packages/core/test/nebul.test.ts b/packages/core/test/nebul.test.ts
index 77f76e53f3c..cced07fce57 100644
--- a/packages/core/test/nebul.test.ts
+++ b/packages/core/test/nebul.test.ts
@@ -1,6 +1,7 @@
 import { expect, test } from "bun:test";
 
 import type { ExistingModel } from "../src/sync/index.js";
+import { MissingReasoningOptionsError } from "../src/sync/missing-reasoning-options.js";
 import {
   NebulEntry,
   NebulResponse,
@@ -80,6 +81,14 @@ test("keeps authored effort sets when the host advertises none", () => {
   expect(translated?.model.reasoning_options).toEqual(authored);
 });
 
+test("fails closed when a reasoner advertises no efforts and none are authored", () => {
+  const entry = nebulEntry("zai-org/GLM-5.3", { reasoning_efforts: [] });
+  expect(() => nebul.translateModel(entry, context(undefined))).toThrow(MissingReasoningOptionsError);
+  expect(() =>
+    nebul.translateModel(entry, context({ base_model: "zhipuai/glm-5.3" } as ExistingModel)),
+  ).toThrow(MissingReasoningOptionsError);
+});
+
 test("keeps existing entries when the source pricing or context is temporarily null", () => {
   const existing = {
     base_model: "zhipuai/glm-5.3",
diff --git a/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml b/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
index 19509361dbc..d6f8fe1901d 100644
--- a/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
+++ b/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
@@ -1,12 +1,16 @@
-# Model reasons always-on; Nebul exposes no thinking toggle or budget field (its
-# docs cover reasoning_effort only): https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+# Efforts mirror Nebul's documented host-wide reasoning_effort enum (its docs
+# note the effort-to-budget mapping is model-specific guidance):
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning. Same-surface
+# peers expose a thinking toggle only, with no effort values to intersect.
 base_model = "alibaba/qwen3.5-397b-a17b"
 
-reasoning_options = []
-
 [interleaved]
 field = "reasoning_content"
 
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "max"]
+
 [cost]
 input = 0.63
 output = 3.78
diff --git a/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
index c86aafb7ef4..6459efe54b7 100644
--- a/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
+++ b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
@@ -1,6 +1,6 @@
 # Effort: reasoning_effort (Nebul docs: accepted for reasoning-capable models, ignored
-# only by non-reasoning models). Set is OpenRouter's low|medium|xhigh intersected with
-# Nebul's documented value set; Nebul has no toggle and xhigh is undocumented on it.
+# only by non-reasoning models). Set is OpenRouter's low|medium|xhigh plus Nebul's
+# documented high; xhigh is undocumented on Nebul and no peer advertises max.
 # https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
 base_model = "alibaba/qwen3.8-27b"
 
@@ -9,7 +9,7 @@ field = "reasoning_content"
 
 [[reasoning_options]]
 type = "effort"
-values = ["low", "medium"]
+values = ["low", "medium", "high"]
 
 [cost]
 input = 0.21
diff --git a/providers/nebul/models/google/gemma-4-31B-it.toml b/providers/nebul/models/google/gemma-4-31B-it.toml
index 756e52c6fa5..f7ca96e731a 100644
--- a/providers/nebul/models/google/gemma-4-31B-it.toml
+++ b/providers/nebul/models/google/gemma-4-31B-it.toml
@@ -1,12 +1,16 @@
-# Model reasons always-on; Nebul exposes no caller control for it (its docs cover
-# reasoning_effort only): https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+# Efforts mirror Nebul's documented host-wide reasoning_effort enum (its docs
+# note the effort-to-budget mapping is model-specific guidance):
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning. Same-surface
+# peers expose a toggle only, with no effort values to intersect.
 # Traces arrive inline as <thinking> inside the content (Nebul docs), so interleaved = true
 base_model = "google/gemma-4-31b-it"
 
-reasoning_options = []
-
 interleaved = true
 
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "max"]
+
 [cost]
 input = 0.42
 output = 1.05
diff --git a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
index c3018447d6d..4c22765a143 100644
--- a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
@@ -1,12 +1,16 @@
-# Hybrid reasoning; Nebul exposes no caller control for it (its docs cover
-# reasoning_effort only): https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+# Efforts mirror Nebul's documented host-wide reasoning_effort enum (its docs
+# note the effort-to-budget mapping is model-specific guidance):
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning. Same-surface
+# peers expose a toggle only, with no effort values to intersect.
 base_model = "nvidia/nemotron-3-nano-30b-a3b"
 
-reasoning_options = []
-
 [interleaved]
 field = "reasoning_content"
 
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "max"]
+
 [cost]
 input = 1
 output = 3
diff --git a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
index abead25f50b..7fca9c970dd 100644
--- a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
@@ -1,12 +1,16 @@
-# Hybrid reasoning; Nebul exposes no caller control for it (its docs cover
-# reasoning_effort only): https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+# Efforts mirror Nebul's documented host-wide reasoning_effort enum (its docs
+# note the effort-to-budget mapping is model-specific guidance):
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning. Same-surface
+# peers expose a toggle only, with no effort values to intersect.
 base_model = "nvidia/nemotron-3-super-120b-a12b"
 
-reasoning_options = []
-
 [interleaved]
 field = "reasoning_content"
 
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "max"]
+
 [cost]
 input = 0.32
 output = 0.69
diff --git a/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml b/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
index 94094ee7c95..5c9dad1974a 100644
--- a/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
+++ b/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
@@ -1,12 +1,16 @@
-# Hybrid reasoning; Nebul exposes no caller control for it (its docs cover
-# reasoning_effort only): https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+# Efforts mirror Nebul's documented host-wide reasoning_effort enum (its docs
+# note the effort-to-budget mapping is model-specific guidance):
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning. Same-surface
+# peers expose a toggle only, with no effort values to intersect.
 base_model = "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"
 
-reasoning_options = []
-
 [interleaved]
 field = "reasoning_content"
 
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "max"]
+
 [cost]
 input = 0.16
 output = 1.05

From 28b97e389b9cd2eaa0fd77f6e32dd8c76316902b Mon Sep 17 00:00:00 2001
From: Wynand Huizinga <wynand.huizinga@nebul.com>
Date: Fri, 11 Sep 2026 20:23:31 +0200
Subject: [PATCH 13/13] fix(nebul): restrict reasoning_options to
 /model/info-advertised efforts

---
 .../nebul/models/Qwen/Qwen3.5-397B-A17B.toml      | 15 +++++++--------
 providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml  | 15 +++++++--------
 providers/nebul/models/google/gemma-4-31B-it.toml | 15 +++++++--------
 .../NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml      | 15 +++++++--------
 .../NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml   | 15 +++++++--------
 ...emotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml | 15 +++++++--------
 providers/nebul/models/zai-org/GLM-5.3.toml       | 12 +++++++-----
 7 files changed, 49 insertions(+), 53 deletions(-)

diff --git a/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml b/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
index d6f8fe1901d..009e54c19fd 100644
--- a/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
+++ b/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
@@ -1,16 +1,15 @@
-# Efforts mirror Nebul's documented host-wide reasoning_effort enum (its docs
-# note the effort-to-budget mapping is model-specific guidance):
-# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning. Same-surface
-# peers expose a thinking toggle only, with no effort values to intersect.
+# Reasoning control per Nebul's docs: GET /v1/model/info `reasoning_efforts` is the
+# per-model source of truth for "the reasoning_effort values each model meaningfully
+# accepts" (https://docs.nebul.io/docs/inference-api/models/model-catalog). It
+# advertises none for this model, and the Chat Completions API documents no on/off
+# toggle (https://docs.nebul.io/docs/inference-api/models/chat-completions), so this
+# host exposes no caller control here.
 base_model = "alibaba/qwen3.5-397b-a17b"
+reasoning_options = []
 
 [interleaved]
 field = "reasoning_content"
 
-[[reasoning_options]]
-type = "effort"
-values = ["low", "medium", "high", "max"]
-
 [cost]
 input = 0.63
 output = 3.78
diff --git a/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
index 6459efe54b7..e8efa43fbf8 100644
--- a/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
+++ b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
@@ -1,16 +1,15 @@
-# Effort: reasoning_effort (Nebul docs: accepted for reasoning-capable models, ignored
-# only by non-reasoning models). Set is OpenRouter's low|medium|xhigh plus Nebul's
-# documented high; xhigh is undocumented on Nebul and no peer advertises max.
-# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+# Reasoning control per Nebul's docs: GET /v1/model/info `reasoning_efforts` is the
+# per-model source of truth for "the reasoning_effort values each model meaningfully
+# accepts" (https://docs.nebul.io/docs/inference-api/models/model-catalog). It
+# advertises none for this model, and the Chat Completions API documents no on/off
+# toggle (https://docs.nebul.io/docs/inference-api/models/chat-completions), so this
+# host exposes no caller control here.
 base_model = "alibaba/qwen3.8-27b"
+reasoning_options = []
 
 [interleaved]
 field = "reasoning_content"
 
-[[reasoning_options]]
-type = "effort"
-values = ["low", "medium", "high"]
-
 [cost]
 input = 0.21
 output = 0.74
diff --git a/providers/nebul/models/google/gemma-4-31B-it.toml b/providers/nebul/models/google/gemma-4-31B-it.toml
index f7ca96e731a..5accf868656 100644
--- a/providers/nebul/models/google/gemma-4-31B-it.toml
+++ b/providers/nebul/models/google/gemma-4-31B-it.toml
@@ -1,16 +1,15 @@
-# Efforts mirror Nebul's documented host-wide reasoning_effort enum (its docs
-# note the effort-to-budget mapping is model-specific guidance):
-# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning. Same-surface
-# peers expose a toggle only, with no effort values to intersect.
+# Reasoning control per Nebul's docs: GET /v1/model/info `reasoning_efforts` is the
+# per-model source of truth for "the reasoning_effort values each model meaningfully
+# accepts" (https://docs.nebul.io/docs/inference-api/models/model-catalog). It
+# advertises none for this model, and the Chat Completions API documents no on/off
+# toggle (https://docs.nebul.io/docs/inference-api/models/chat-completions), so this
+# host exposes no caller control here.
 # Traces arrive inline as <thinking> inside the content (Nebul docs), so interleaved = true
 base_model = "google/gemma-4-31b-it"
+reasoning_options = []
 
 interleaved = true
 
-[[reasoning_options]]
-type = "effort"
-values = ["low", "medium", "high", "max"]
-
 [cost]
 input = 0.42
 output = 1.05
diff --git a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
index 4c22765a143..0b45ea1bb92 100644
--- a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
@@ -1,16 +1,15 @@
-# Efforts mirror Nebul's documented host-wide reasoning_effort enum (its docs
-# note the effort-to-budget mapping is model-specific guidance):
-# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning. Same-surface
-# peers expose a toggle only, with no effort values to intersect.
+# Reasoning control per Nebul's docs: GET /v1/model/info `reasoning_efforts` is the
+# per-model source of truth for "the reasoning_effort values each model meaningfully
+# accepts" (https://docs.nebul.io/docs/inference-api/models/model-catalog). It
+# advertises none for this model, and the Chat Completions API documents no on/off
+# toggle (https://docs.nebul.io/docs/inference-api/models/chat-completions), so this
+# host exposes no caller control here.
 base_model = "nvidia/nemotron-3-nano-30b-a3b"
+reasoning_options = []
 
 [interleaved]
 field = "reasoning_content"
 
-[[reasoning_options]]
-type = "effort"
-values = ["low", "medium", "high", "max"]
-
 [cost]
 input = 1
 output = 3
diff --git a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
index 7fca9c970dd..e8bd3c054dc 100644
--- a/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
@@ -1,16 +1,15 @@
-# Efforts mirror Nebul's documented host-wide reasoning_effort enum (its docs
-# note the effort-to-budget mapping is model-specific guidance):
-# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning. Same-surface
-# peers expose a toggle only, with no effort values to intersect.
+# Reasoning control per Nebul's docs: GET /v1/model/info `reasoning_efforts` is the
+# per-model source of truth for "the reasoning_effort values each model meaningfully
+# accepts" (https://docs.nebul.io/docs/inference-api/models/model-catalog). It
+# advertises none for this model, and the Chat Completions API documents no on/off
+# toggle (https://docs.nebul.io/docs/inference-api/models/chat-completions), so this
+# host exposes no caller control here.
 base_model = "nvidia/nemotron-3-super-120b-a12b"
+reasoning_options = []
 
 [interleaved]
 field = "reasoning_content"
 
-[[reasoning_options]]
-type = "effort"
-values = ["low", "medium", "high", "max"]
-
 [cost]
 input = 0.32
 output = 0.69
diff --git a/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml b/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
index 5c9dad1974a..943ff62c877 100644
--- a/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
+++ b/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
@@ -1,16 +1,15 @@
-# Efforts mirror Nebul's documented host-wide reasoning_effort enum (its docs
-# note the effort-to-budget mapping is model-specific guidance):
-# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning. Same-surface
-# peers expose a toggle only, with no effort values to intersect.
+# Reasoning control per Nebul's docs: GET /v1/model/info `reasoning_efforts` is the
+# per-model source of truth for "the reasoning_effort values each model meaningfully
+# accepts" (https://docs.nebul.io/docs/inference-api/models/model-catalog). It
+# advertises none for this model, and the Chat Completions API documents no on/off
+# toggle (https://docs.nebul.io/docs/inference-api/models/chat-completions), so this
+# host exposes no caller control here.
 base_model = "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"
+reasoning_options = []
 
 [interleaved]
 field = "reasoning_content"
 
-[[reasoning_options]]
-type = "effort"
-values = ["low", "medium", "high", "max"]
-
 [cost]
 input = 0.16
 output = 1.05
diff --git a/providers/nebul/models/zai-org/GLM-5.3.toml b/providers/nebul/models/zai-org/GLM-5.3.toml
index 0a964c0dda7..09dfa0df440 100644
--- a/providers/nebul/models/zai-org/GLM-5.3.toml
+++ b/providers/nebul/models/zai-org/GLM-5.3.toml
@@ -1,13 +1,15 @@
-# GLM-5.3 always reasons; efforts low|high|max per zhipuai's official entry
+# Reasoning control per Nebul's docs: GET /v1/model/info `reasoning_efforts` is the
+# per-model source of truth for "the reasoning_effort values each model meaningfully
+# accepts" (https://docs.nebul.io/docs/inference-api/models/model-catalog). It
+# advertises an empty list for this model, and the Chat Completions API documents no
+# on/off toggle (https://docs.nebul.io/docs/inference-api/models/chat-completions),
+# so this host exposes no caller control here.
 # Traces arrive in message.reasoning (Nebul docs), so interleaved = true
 base_model = "zhipuai/glm-5.3"
+reasoning_options = []
 
 interleaved = true
 
-[[reasoning_options]]
-type = "effort"
-values = ["low", "high", "max"]
-
 [cost]
 input = 1.47
 output = 4.62