feat: add Cohere, Perplexity, and Meta providers

[brunobuddy]

💭 Why

Several current providers and models people look up were missing. Cohere, Perplexity, and Meta had logos but no entries, Alibaba had only one Qwen model, and a few current OpenAI reasoning models plus MiniMax M3 weren't covered.

✨ What changed

• New provider Cohere: 8 Command models (A+, A, A-Reasoning, A-Vision, A-Translate, R7B, R+, R).
• New provider Perplexity: 4 Sonar models (sonar, sonar-pro, sonar-reasoning-pro, sonar-deep-research).
• New provider Meta: 4 Llama API models, plus a meta.svg logo.
• Alibaba Qwen: 1 to 8 models (qwen3-max, qwen3.5-flash, qwen-plus, qwen-flash, qwen3-coder-plus, qwen3-coder-flash, qwq-plus).
• OpenAI: gpt-5.3-codex, gpt-5.4-pro, gpt-5.4-nano, gpt-5.5-pro, o3-pro, plus gpt-5.4-pro / gpt-5.5-pro subscription variants.
• MiniMax M3 (api_key + subscription).
• Magistral medium/small: add the missing prompt_mode reasoning control.

📝 Notes

Catalog grows 142 to 174 models. validate, typecheck, lint, tests, guard:params, and build pass locally. Every parameter is verified against the provider's official API reference. A few choices worth a reviewer's eye:

• Cohere frequency/presence penalties use a 0 to 1 range (not OpenAI's -2..2), per Cohere's reference.
• gpt-5.4-pro effort is documented (medium/high/xhigh). gpt-5.5-pro and o3-pro do not publish the enum, so they follow their pro / o3 siblings.
• Meta does not publish numeric ranges for Llama's sampling knobs, so those are listed without ranges instead of guessing.
• Qwen thinking reuses the existing extra_body.chat_template_kwargs.enable_thinking path for consistency with the qwen3.5 entry.
• MiniMax M3 sampling ranges mirror the M2.x entries (0.01 to 1). M3 docs may allow temperature up to 2.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
modelparams.dev	Ready	Preview	Jun 2, 2026 4:15pm