feat: add Gemma open weights, GLM-5.2, Qwen 3.7, Kimi K2.7

[brunobuddy]

💭 Why

Open-weight models stay useful after a provider drops them from its hosted API, since the weights still run locally. The catalog was also missing several current flagships.

✨ What changed

• Add Gemma 3 (27b/12b/4b/1b), Gemma 3n (E4B/E2B), and Gemma 4 (12B/E4B/E2B) as local open-weight entries with model-card params.
• Add z-ai/glm-5.2 (API + Coding Plan): the glm-5.1 set plus the new reasoning_effort enum (default max, thinking-only).
• Add alibaba/qwen3.7-max, qwen3.7-plus, qwen3.6-flash with extra_body.enable_thinking and extra_body.thinking_budget.
• Add moonshot/kimi-k2.7-code (+-highspeed): temperature/top_p are pinned by the API and thinking is always on, so only max tokens and response format are listed.
• Add google/gemini-3.1-flash-lite API-key entry (GA since May 7); fix the subscription entry's thinking default (docs table says minimal, not high).
• Add GLM-5.2 display label; commit codegen output.

Type of change

• Add a model (new YAML file under models/)
• Add or update parameters on an existing model
• Fix incorrect data (default, range, values, applicability)

Source

• Gemma: Gemma 4 model card (standardized temperature 1.0, top_p 0.95, top_k 64; context per size), Gemma 3 model card, Gemma 3n model card, shipped generation_config.json.
• GLM-5.2: chat completion reference, parameter concepts (output default 65536, max 131072), coding plan.
• Qwen: chat completions reference (max_completion_tokens for new integrations, extra_body.enable_thinking, top_k default 20), deep thinking (all three hybrid, thinking on by default).
• Kimi: chat API, models overview (pinned sampling values error if changed), K2.7 Code quickstart (default 32768 output).
• Gemini: gemini-3.1-flash-lite model page, Gemini 3 thinking-level table (minimal is the flash-lite default).

📝 Notes

• Gemma 3/3n are no longer served on the Gemini API, so these entries use bare local-serving paths (max_completion_tokens, temperature, top_p, top_k) like the Meta entries, not generationConfig.*. Model ids follow Hugging Face casing.
• New Qwen entries use extra_body.enable_thinking, the path the current hosted chat-completions reference documents (older entries use chat_template_kwargs).
• Stacked on feat: add gpt-oss family + Thinking Machines Inkling #135; merge that one first.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
modelparams.dev	Ready	Preview	Jul 20, 2026 10:21am