feat: add MiniMax as first-class LLM provider (M3 default)

[octo-patch]

Summary

• Add MiniMax AI as a first-class LLM provider alongside OpenAI, Cerebras, and Azure OpenAI
• MiniMax is detected via MINIMAX_API_KEY env var and uses its OpenAI-compatible API at https://api.minimax.io/v1
• Default model is MiniMax-M3 (512K context, max output 128K, image input support); MiniMax-M2.7 and MiniMax-M2.7-highspeed remain available as alternatives
• Temperature values are auto-clamped to MiniMax's valid range (0, 1] to prevent API errors

Changes

Provider Detection (optillm/server.py)

• Added MINIMAX_API_KEY check in get_config() between Cerebras and OpenAI in priority order
• Creates an OpenAI client with MiniMax base URL (https://api.minimax.io/v1) by default
• Supports custom base_url override via --base-url flag
• Properly passes SSL/httpx configuration to the client

Temperature Clamping (optillm/server.py)

• Clamps temperature to 0.01 when <= 0 (MiniMax requires strictly positive temperature)
• Clamps temperature to 1.0 when > 1.0
• Only applies when MINIMAX_API_KEY is set; other providers are unaffected

Documentation (README.md)

• Added MiniMax row to the provider table with API key, link, and notes (M3 listed as default; M2.7 and M2.7-highspeed as alternatives)

Tests

• 17 unit tests (tests/test_minimax_provider.py): provider detection, base URL, SSL, priority ordering, temperature clamping
• 3 integration tests (tests/test_minimax_integration.py): basic completion, temperature boundary, streaming (requires live API key; default model is MiniMax-M3)

Provider Priority Order

1. OPTILLM_API_KEY - Local inference
2. CEREBRAS_API_KEY - Cerebras
3. MINIMAX_API_KEY - MiniMax (new)
4. OPENAI_API_KEY - OpenAI
5. AZURE_OPENAI_API_KEY - Azure OpenAI
6. LiteLLM fallback

Usage

export MINIMAX_API_KEY="your-key"
optillm

# Then use with any approach:
# model="moa-MiniMax-M3" or model="bon-MiniMax-M3"
# Older variants still work: model="moa-MiniMax-M2.7"

Supported Models

Model	Notes
MiniMax-M3 (default)	512K context, up to 128K output, image input
MiniMax-M2.7	Previous generation
MiniMax-M2.7-highspeed	Previous generation, low-latency variant

Test plan

• All 17 unit tests pass
• All 3 integration tests pass with live MiniMax API (M3)
• Verify no regressions in existing provider detection (OpenAI, Cerebras, Azure)

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}