[roykollensvendsen]

Adds Verathos Decide (Bittensor subnet 96), which serves open-weight decision models on POST /v1/systemone in the System One request and answer format. It has no chat completions endpoint. @aizorr0, this replaces #7243, which closed as stale after Verathos moved from chat models to decision models; please say if anything here doesn't match how you serve them.

Changes

• providers/verathos/: provider.toml (VERATHOS_API_KEY, https://api.verathos.ai/v1), logo.svg (Verathos' own favicon with the fill set to currentColor) and four model files. provider.toml sets opencodeai = "none", since Verathos has no chat API for the provider default to describe. All four are v2 v2_only files with base_model, an [api.systemone] table (with sdk = { aisdk = "@ai-sdk/typesafe-ai", opencodeai = "@opencode/ai/providers/typesafe-ai" }, as AMD, Baseten and Melious set it), cost and the host's real deltas:

Model	Input $/MTok	Context here	Delta from the lab entry
kev-4b	0.015	65,536	none
clef-flash	0.02	65,536	context capped (the lab entry is 262,144), temperature = false (System One has no temperature field, as on Melious); modalities inherited from the lab, as on Melious and OpenRouter
d3-flash	0.02	16,384	text only, context capped
d3-mini	0.015	16,384	text only, context capped

The D3 files narrow input to text and keep output = ["decision"] next to it, because the v2 schema requires both keys in a provider file's [modalities]: without output, bun validate fails on modalities.output: Required.

Output is 0 everywhere: decision models generate no text, and Verathos bills input tokens only.

• New lab entries, since none existed: models/jaredpalmer/kev-4b and models/vllm-sr/d3-mini, both v2 v2_only like d3 and d3-flash from feat(v2): add new models for the 11 small providers #10758, from their Hugging Face model cards. D3 Mini's context of 262,144 is what the vLLM-SR runtime serves (decision_config.json sets max_length = null), the same as d3 and d3-flash after fix(vllm-sr): give the d3 and d3-flash lab entries the runtime's context #10827. d3-flash uses the existing models/vllm-sr/d3-flash.
• packages/core/src/generate.ts: v1 generation leaves out a provider whose models are all v2_only instead of failing with has no models, with a test in generate.test.ts. This is the same change as in feat(decisionmodels): add Decision Models provider with ten decision models #10818; whichever merges first carries it and the other rebases cleanly.

Sources (accessed 2026-10-20)

• Prices, max_state_tokens, input_modalities and weight revisions: GET https://api.verathos.ai/v1/models (public, no key)
• API, auth and billing: https://verathos.ai/docs/api, https://verathos.ai/docs/authentication, https://verathos.ai/docs/models
• Lab facts: https://huggingface.co/jaredpalmer/kev-4b, https://huggingface.co/vllm-sr/d3-flash, https://huggingface.co/vllm-sr/d3-mini. Verathos pins d3 to the commits after tag v3.1.0, from 2026-10-11 to 2026-10-12.

Live check (2026-10-20, with a Verathos key)

• One text request per model returned 200, and usage.input_tokens times the listed price matched the file.
• An image request to clef-flash returned 200 (usage.image_tokens 67). The same image to kev-4b, d3-flash and d3-mini returned 400 invalid_request ("reads text only"). That is why the D3 entries are text only.
• Each model refused a state just over its max_state_tokens with 413 too_large. d3-flash accepted 16,158 tokens of prose, and its real tokenizer, not the documented 4 characters per token, decides the cap.
• Total spend: under $0.001.

Test plan

• bun validate
• bun run test in packages/sdk (25 pass)
• v1 output against upstream/dev: identical (Verathos is v2 only, so v1 does not list it)
• v2 output: four verathos models, type decision, with the prices, limits and modalities above
• bun test test/generate.test.ts in packages/core: the new test passes, and fails without the generator change

🤖 Generated with Claude Code

https://claude.ai/code/session_018k8EL6s5hEtTTMJ8Jjd4Ls

[github-actions]

No actionable findings.

[roykollensvendsen]

A sync module for these files is ready as draft #10666, stacked on this PR. It keeps prices, max_state_tokens and input modalities current from the public /v1/models, never creates or deletes files, and leaves the four files here unchanged against today's list. I'll rebase it onto dev and mark it ready once this merges.

[github-actions]

Action items

• [medium] [possible mistake] models/vllm-sr/d3-mini.toml:28 - Check: Lab limit.context must reflect documented usable context, not raw architecture max. Why: This sets 262_144 from max_position_embeddings because the card states none, while same-family siblings models/vllm-sr/d3.toml and d3-flash.toml use 32_768 from serving docs; claiming 8× larger context for the smaller 4B sibling from a config field is internally inconsistent and risks routing oversized states to the lab limit. Action: Provide a direct source that d3-mini weights serve 262,144 usable tokens, or align the lab limit methodology with the d3/d3-flash siblings.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.

[github-actions]

Action items

• [high] [violation] providers/verathos/models/clef-flash.toml:9 - Check: Provider base_model files must be override-only and never restate identical data. Why: input = ["text", "image/png", "image/jpeg", "image/webp"] is byte-identical to models/cloudflare/clef-flash.toml input, so it is not a delta despite the PR table claiming it as one. Action: Drop the identical input line and keep only real deltas in [modalities].
• [high] [violation] providers/verathos/models/d3-flash.toml:10 - Check: Provider base_model files must keep only real deltas. Why: output = ["decision"] is identical to models/vllm-sr/d3-flash.toml output; [limit]/[modalities] objects deep-merge so input can be overridden alone. Action: Remove output = ["decision"] and keep only input = ["text"].
• [high] [violation] providers/verathos/models/d3-mini.toml:10 - Check: Provider base_model files must keep only real deltas. Why: output = ["decision"] is identical to the new models/vllm-sr/d3-mini.toml output in the same PR; only the text-only input cap is a real delta. Action: Remove output = ["decision"] and keep only input = ["text"].
• [medium] [possible mistake] providers/verathos/models/clef-flash.toml:10 - Check: System One modalities.output must match host behavior and lab/peer evidence. Why: Lab models/cloudflare/clef-flash.toml is output = ["text"] and same-surface System One peers providers/melious/models/clef-flash.toml and providers/openrouter/models/cloudflare/clef-flash.toml keep output = ["text"] with structured_output, while this file flips to output = ["decision"] without cited evidence for the difference. Action: Verify against https://api.verathos.ai/v1/models / https://verathos.ai/docs/api whether Verathos returns decision vs text-structured output on POST /v1/systemone, and either cite it in the leading comment or align to ["text"] like peers.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.