[guillaumegay13]

Automated sync between the Manifest gateway and this catalog (modelparams-agent).

How entries are verified: the local gateway loads this branch's catalog; for every declared param and test value the agent configures it through the gateway's model-params API, sends a request through the proxy pinned to the model, and checks the gateway applied the value upstream and got a 200 from that model. Params that fail are removed from the entry before it is filed.

This PR is rebuilt from main on every run and merged automatically once CI is green.

Added (1)

Entry	Params verified	Dropped	Corrected	Cloned from
nvidia/llama-3.2-90b-vision-instruct.yaml	temperature top_p max_tokens frequency_penalty presence_penalty	—	—	nvidia/llama-3.2-11b-vision-instruct.yaml

Evidence

models/nvidia/llama-3.2-90b-vision-instruct.yaml

• baseline: 200 from nvidia/meta/llama-3.2-90b-vision-instruct (api_key)
• temperature ✅ 0.6, 0, 1
• top_p ✅ 0.95, 1
• max_tokens ✅ 4096, 1, 16384
• frequency_penalty ✅ 0, -2, 2
• presence_penalty ✅ 0, -2, 2

Corrected (3)

Existing entries re-verified through the gateway (daily rotation). A declared bound was refused by the vendor and one step inside was accepted; or enum values were refused by the vendor and the corrected list (refused values removed, values the vendor lists as supported added) was served to the gateway on a draft catalog and every value passed.

• models/fireworks/inkling.yaml
  • top_p: declared min 0 refused (HTTP 400: top_p must be in (0, 1], got 0.0.); 0.01 accepted → min corrected to 0.01
• models/google/gemini-3.7-flash.yaml
  • generationConfig.thinkingConfig.thinkingLevel: values [minimal, low, medium, high] → [low, medium, high]; default minimal → low
  • vendor: "Thinking level MINIMAL is not supported for this model. Please retry with other thinking level."
  • proof through the gateway (draft catalog with the corrected entry): low ✅ medium ✅ high ✅
  • note: default minimal refused and the vendor names no default; set to low (nearest remaining value in the declared order) — a maintainer may prefer the vendor's documented default
  • other params re-verified on the corrected entry: generationConfig.maxOutputTokens generationConfig.temperature generationConfig.topP generationConfig.topK generationConfig.thinkingConfig.includeThoughts generationConfig.responseMimeType ✅
• models/groq/gpt-oss-20b.yaml
  • reasoning_format: values [hidden, raw, parsed] → [hidden, parsed]; default undefined (unchanged)
  • vendor: "reasoning_format=raw is not supported with this model"
  • proof through the gateway (draft catalog with the corrected entry): hidden ✅ parsed ✅
  • other params re-verified on the corrected entry: max_completion_tokens temperature top_p frequency_penalty presence_penalty reasoning_effort response_format.type ✅

Removing an enum value is not a parameter removal for guard:params (it compares param paths per model, not values), and the removed values are refused by the vendor anyway.

Needs a human: params now refused (5)

Re-verification found these declared params refused. Not removed — the param guard blocks removals and dropping a param is a maintainer's call.

• models/mistral/magistral-medium-latest.yaml prompt_mode: prompt_mode="reasoning" → HTTP 400: Reasoning prompt mode is not enabled for this model
• models/mistral/magistral-medium-latest.yaml response_format.type: response_format.type="json_schema" → HTTP 400: Schema response format type requires a json_schema
• models/mistral/mistral-medium-latest.yaml response_format.type: response_format.type="json_schema" → HTTP 400: Schema response format type requires a json_schema
• models/moonshot/kimi-k2.7-code.yaml response_format.type: response_format.type="json_schema" → HTTP 400: Invalid request: missing required parameter: 'response_format.json_schema'
• models/nvidia/gpt-oss-20b.yaml seed: seed=18446744073709552000 → HTTP 400: 1 validation error: {'type': 'less_than_equal', 'loc': ('body', 'seed'), 'msg': 'Input should be less than or equal to 9223372036854775807', 'input': 1844674407

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
modelparams.dev	Ready	Preview	Oct 2, 2026 5:43am UTC