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/packages/core/src/sync/index.ts b/packages/core/src/sync/index.ts
index e6c30874235..911ae65fb89 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 { ollamaCloud } from "./providers/ollama-cloud.js";
 import { openai } from "./providers/openai.js";
@@ -153,6 +154,7 @@ export const providers: {
   "llmgateway-providers": SyncProvider<any>;
   "merge-gateway": SyncProvider<any>;
   meta: SyncProvider<any>;
+  nebul: SyncProvider<any>;
   "nano-gpt": SyncProvider<any>;
   ofox: SyncProvider<any>;
   "ollama-cloud": SyncProvider<any>;
@@ -189,6 +191,7 @@ export const providers: {
   "llmgateway-providers": llmgatewayProviders,
   "merge-gateway": mergeGateway,
   meta,
+  nebul,
   "nano-gpt": nanoGpt,
   ofox,
   "ollama-cloud": ollamaCloud,
@@ -222,7 +225,7 @@ export const groups = {
     "vercel",
   ],
   cloudflare: ["cloudflare-ai-gateway", "cloudflare-workers-ai"],
-  direct: ["ambient", "anthropic", "baseten", "chutes", "cortecs", "deepinfra", "digitalocean", "github-copilot", "google", "hyper", "meta", "ollama-cloud", "openai", "ovhcloud", "pioneer", "tinfoil", "venice", "wandb", "xai"],
+  direct: ["ambient", "anthropic", "baseten", "chutes", "cortecs", "deepinfra", "digitalocean", "github-copilot", "google", "hyper", "meta", "nebul", "ollama-cloud", "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..b3010f0a5c1
--- /dev/null
+++ b/packages/core/src/sync/providers/nebul.ts
@@ -0,0 +1,212 @@
+import { existsSync, readdirSync } from "node:fs";
+import path from "node:path";
+import { z } from "zod";
+
+import type { ExistingModel, SyncProvider, SyncedModel } from "../index.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");
+
+// 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) {
+    const data = NebulResponse.parse(raw).data;
+    // An empty catalog is an upstream fault; syncing it would delete every
+    // local model file via the delete-missing pass, so fail loudly instead.
+    if (data.length === 0) {
+      throw new Error("Nebul returned an empty model catalog");
+    }
+    // Same failure mode if the response shape drifts and no entry matches the
+    // chat-model filter anymore (e.g. renamed model_type/mode values).
+    if (!data.some(isCatalogChatModel)) {
+      throw new Error("Nebul returned no usable chat models");
+    }
+    return 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. Whole-catalog faults
+  // fail closed in parseModels before any file is written or deleted.
+  translateModel(entry, context) {
+    if (!isCatalogChatModel(entry)) return undefined;
+    const id = entry.model_name;
+    const info = entry.model_info;
+    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;
+    // 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),
+      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.
+  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) => `\`${id}\``).join(", "),
+    ];
+  },
+} satisfies SyncProvider<NebulEntry>;
+
+function isCatalogChatModel(entry: NebulEntry): boolean {
+  const info = entry.model_info;
+  return info.model_type === "llm" && info.mode === "chat"
+    && entry.model_name !== PING_MODEL && !DEPRECATED.has(entry.model_name) && !DENYLIST.test(entry.model_name);
+}
+
+// Nebul documents exactly one reasoning control: reasoning_effort. When the
+// host advertises efforts, write that effort entry and nothing else —
+// lab-style toggles or budgets are not supported on this API. When it
+// advertises none, keep the authored options; a reasoner with neither is
+// rejected above so no empty options entry is ever synced.
+function buildReasoningOptions(entry: NebulEntry, existing: ExistingModel | undefined) {
+  const efforts = entry.model_info.reasoning_efforts ?? [];
+  if (efforts.length === 0) return existing?.reasoning_options;
+  return [{ type: "effort" as const, values: efforts }];
+}
+
+function resolveBaseModel(servedID: string, huggingfaceID: string | undefined): string | undefined {
+  return baseModelCandidates(servedID, huggingfaceID).find(canonicalExists);
+}
+
+// existsSync is case-insensitive on Windows/macOS; verify the real on-disk filename case
+// so the resolved base_model matches the canonical metadata exactly (and CI on Linux).
+function canonicalExists(candidate: string): boolean {
+  const file = path.join(MODELS_DIR, `${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..44610463e72
--- /dev/null
+++ b/packages/core/test/nebul.test.ts
@@ -0,0 +1,196 @@
+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,
+  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("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: "effort", values: ["low", "high", "max"] }]);
+});
+
+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 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("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",
+    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"],
+    ["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("fails closed on an empty catalog so sync cannot delete every local file", () => {
+  expect(() => nebul.parseModels({ data: [] })).toThrow("Nebul returned an empty model catalog");
+});
+
+test("fails closed when no entry matches the chat-model filter", () => {
+  expect(() =>
+    nebul.parseModels({
+      data: [
+        { model_name: "Some/Embedding", model_info: { mode: null, model_type: "embedding" } },
+        { model_name: "Some/Reranker", model_info: { mode: null, model_type: "rerank" } },
+      ],
+    }),
+  ).toThrow("Nebul returned no usable chat models");
+});
+
+test("parseModels keeps chat entries alongside filtered serving artifacts", () => {
+  const parsed = nebul.parseModels({
+    data: [
+      { model_name: "Some/Embedding", model_info: { mode: null, model_type: "embedding" } },
+      { model_name: "Some/Chat", model_info: { mode: "chat", model_type: "llm" } },
+    ],
+  });
+  expect(parsed).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..07d1d5954e0
--- /dev/null
+++ b/providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml
@@ -0,0 +1,17 @@
+# 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.
+# Trace field: reasoning_content — that page names Qwen3 thinking variants explicitly.
+base_model = "alibaba/qwen3.5-397b-a17b"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[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..f00d0f1fcd5
--- /dev/null
+++ b/providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml
@@ -0,0 +1,17 @@
+# 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.
+# Trace field: reasoning_content — that page names Qwen3 thinking variants explicitly.
+base_model = "alibaba/qwen3.8-27b"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[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..54b41c88bea
--- /dev/null
+++ b/providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml
@@ -0,0 +1,19 @@
+# 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
+# Trace field: reasoning_content per that same page (Nebul's channel for OpenAI-style reasoners).
+base_model = "deepseek/deepseek-v4.1-flash"
+
+[interleaved]
+field = "reasoning_content"
+
+[[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..780dcf4dccd
--- /dev/null
+++ b/providers/nebul/models/google/gemma-4-31B-it.toml
@@ -0,0 +1,16 @@
+# 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, so interleaved = true
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+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
new file mode 100644
index 00000000000..91a7725a533
--- /dev/null
+++ b/providers/nebul/models/meta-models/muse-glimmer-30b.toml
@@ -0,0 +1,18 @@
+# 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.
+# Trace field: reasoning_content per https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+base_model = "meta/muse-glimmer-30b"
+
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[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..027d87b6fba
--- /dev/null
+++ b/providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml
@@ -0,0 +1,16 @@
+# 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
+# Trace field: reasoning_content per https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+base_model = "mistral/mistral-medium-2604"
+
+[interleaved]
+field = "reasoning_content"
+
+[[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..2c90aaf6ce6
--- /dev/null
+++ b/providers/nebul/models/moonshotai/Kimi-K3.toml
@@ -0,0 +1,19 @@
+# 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 — the live response also marks this served ID with
+# supports_reasoning = false — 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.
+# Trace field: reasoning_content per https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+base_model = "moonshotai/kimi-k3"
+
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[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..66a123d5534
--- /dev/null
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml
@@ -0,0 +1,17 @@
+# 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.
+# Trace field: reasoning_content — that page names the Nemotron family explicitly.
+base_model = "nvidia/nemotron-3-nano-30b-a3b"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[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..ea46772ad61
--- /dev/null
+++ b/providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml
@@ -0,0 +1,20 @@
+# 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.
+# Trace field: reasoning_content — that page names the Nemotron family explicitly.
+base_model = "nvidia/nemotron-3-super-120b-a12b"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[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..3f6caaa49df
--- /dev/null
+++ b/providers/nebul/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.toml
@@ -0,0 +1,20 @@
+# 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.
+# Trace field: reasoning_content — that page names the Nemotron family explicitly.
+base_model = "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[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..c4abddea461
--- /dev/null
+++ b/providers/nebul/models/openai/gpt-oss-120b.toml
@@ -0,0 +1,16 @@
+# Efforts: reasoning_effort = low|medium|high — the model's documented Chat Completions control
+# Trace field: reasoning_content for OpenAI-style models per
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+base_model = "openai/gpt-oss-120b"
+
+[interleaved]
+field = "reasoning_content"
+
+[[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..f834a0f0b81
--- /dev/null
+++ b/providers/nebul/models/zai-org/GLM-5.3-Flash.toml
@@ -0,0 +1,20 @@
+# Efforts: reasoning_effort = low|high|max on POST /v1/chat/completions (verified live 2026-09-09)
+# Traces arrive in message.reasoning (GLM-5.x family), so interleaved = true
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+# limit.context = 1_048_572 is the live GET /v1/model/info max_input_tokens for this
+# exact ID (re-checked 2026-09-13); sibling GLM-5.3 is 1_048_576 — the 4-token
+# difference between the two IDs is real, not a typo.
+base_model = "zhipuai/glm-5.3-flash"
+
+interleaved = true
+
+[[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..fe18dc8e5a6
--- /dev/null
+++ b/providers/nebul/models/zai-org/GLM-5.3.toml
@@ -0,0 +1,23 @@
+# 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). The live
+# catalog (https://api.inference.nebul.io/model/info) advertises an empty list for
+# this exact ID — unlike sibling zai-org/GLM-5.3-Flash, which advertises
+# ["low","high","max"], so the differing values between the two IDs are intentional
+# — 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 (GLM-5.x family), so interleaved = true
+# https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
+base_model = "zhipuai/glm-5.3"
+reasoning_options = []
+
+interleaved = true
+
+[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"
diff --git a/sync.md b/sync.md
index 06ceb1b2a42..acc43b5db69 100644
--- a/sync.md
+++ b/sync.md
@@ -279,6 +279,17 @@ xAI is implemented in `packages/core/src/sync/providers/xai.ts`.
 - New documented models open deduped missing-model issues for manual authoring (`skipCreates`); local models absent from the docs are retained (`deleteMissing: false`). Image generation, transcription, and self-hosted models are outside this sync's scope.
 - Missing tables, unknown pricing tiers, invalid prices/limits, and duplicate model rows fail before writing. Documentation format changes require updating the parser, not guessing defaults.
 
+## Nebul Notes
+
+Nebul is implemented in `packages/core/src/sync/providers/nebul.ts`.
+
+- Source endpoint: `https://api.inference.nebul.io/model/info` (the `/v1` alias also works). No authentication; with an API key the response would be project-scoped, so the sync intentionally calls it unauthenticated for the full public catalog.
+- The endpoint is treated as the authoritative catalog: entries removed server-side are removed locally, and new resolvable chat models are created with `base_model` overrides only.
+- Whole-catalog faults fail closed: an empty response, or one where nothing matches the chat-model filter (`model_type: "llm"`, `mode: "chat"`), throws in `parseModels` before any file is written or deleted — mirroring the LLM Gateway guards.
+- Per-model robustness is the inverse: an existing entry survives transient null pricing, a missing context limit, or a served alias that no longer resolves to lab metadata, keeping its authored `base_model`/`cost`/`limit`; only brand-new models require a fully-priced, resolvable source entry. Embeddings, rerankers, OCR/guard models, the ping health check, and server-side-superseded GLM-5.1/5.2 IDs skip silently.
+- `reasoning_options` come from the endpoint's per-model `reasoning_efforts` list, Nebul's only documented reasoning control. A reasoner with neither advertised efforts nor authored controls fails sync for manual authoring rather than writing an empty control set.
+- `interleaved`, `status`, and other fields the endpoint does not expose are preserved from existing files; the reasoning trace channel (`reasoning_content` vs `message.reasoning` vs inline `<thinking>`) is family-specific per Nebul's docs.
+
 ## OVHcloud Notes
 
 OVHcloud AI Endpoints is implemented in `packages/core/src/sync/providers/ovhcloud.ts`.