modelsettings

Model Settings Registry

This package applies model tuning settings independent of serving provider.

Goal

Use a bounded model list (OpenRouter catalog), then resolve parameter defaults by precedence:

1. explicit request override
2. creator recommendation (exact model)
3. creator recommendation (model family)
4. third-party fallback (OpenRouter model defaults)
5. provider config fallback

If a parameter is undefined by creator docs, it remains unset (null / omitted).

Files

• openrouter_model_settings.json: bounded model catalog snapshot from OpenRouter.
• creator_recommendations.json: curated creator-sourced recommendations and family rules.
• settings.go: resolver, normalization, and precedence merge.

Priority open-model families

Current creator-priority profiles focus on:

• DeepSeek (deepseek-*, plus deepseek-r1* overrides)
• MiniMax (minimax-m2*)
• Qwen (qwen3*, qwen3.5*, qwen3.6*, qwen3.8*, qwen2.5*, qwen2.5-coder*)
• ZAI GLM (exact creator-backed rules currently for glm-4.6 and glm-4-9b-chat)

Gemma currently has no accessible creator-published sampling defaults in this environment; values remain unset unless explicit overrides are provided.

Mode-aware parameters

Some models (Qwen3.6, Qwen3.8) publish different creator-recommended sampling for thinking vs. non-thinking (instruct) mode. A profile can declare an optional instruct_parameters block. ResolveModelSettingsForMode(model, instruct) returns instruct_parameters when instruct is true and the profile defines the block, otherwise the default parameters. The request builder passes disableThinking as instruct, so non-thinking requests get the instruct recommendation. ResolveModelSettings(model) is unchanged and resolves the default (thinking) mode.

Normalization

Model names are normalized to base model IDs:

• strips provider prefix (openai/gpt-oss-20b -> gpt-oss-20b)
• strips OpenRouter variants (:free, :nitro, etc.)
• strips common quant suffixes (_Q4, -int8, -gguf, etc.)

This makes variants like gpt-oss-20b_Q4 resolve to the same recommendation profile as openai/gpt-oss-20b.

Refresh OpenRouter catalog

curl -s https://openrouter.ai/api/v1/models > /tmp/openrouter_models.json
jq -c '{
  generated_at: (now|todate),
  source: "https://openrouter.ai/api/v1/models",
  model_count: (.data|length),
  models: [.data[] | {
    id: .id,
    slug: (.id|split("/")[1]),
    supported_parameters: .supported_parameters,
    default_parameters: {
      temperature: .default_parameters.temperature,
      top_p: .default_parameters.top_p,
      frequency_penalty: .default_parameters.frequency_penalty
    }
  }]
}' /tmp/openrouter_models.json > pkg/model_settings/openrouter_model_settings.json