MiniMax M3 API Guide

MiniMax-M3 is registered as a first-class model in this deployment. It is available in the hosted Chat model selector, the OpenAI-compatible API, and the Anthropic-compatible API without changing endpoint paths.

Official MiniMax references:

• MiniMax M3 model page
• MiniMax M3 release report
• Anthropic SDK reference
• OpenAI SDK reference

Model Identity

• Model ID: MiniMax-M3
• Upstream model ID: MiniMax-M3
• Billing model ID: MiniMax-M3
• Aliases: MiniMax M3, codex-MiniMax-M3
• Official context window: 1,000,000 tokens
• Input modalities: text, image, video
• Channels: hosted Chat, OpenAI-compatible API, Anthropic-compatible API

Existing channel defaults remain unchanged. Select M3 explicitly with the model field.

OpenAI-Compatible Request

curl https://minimax-m2.com/api/v1/chat/completions \
  -H "content-type: application/json" \
  -H "authorization: Bearer $MINIMAX_API_KEY" \
  -d '{
    "model": "MiniMax-M3",
    "messages": [
      {"role": "user", "content": "Review this architecture."}
    ],
    "max_completion_tokens": 2048,
    "reasoning_split": true
  }'

For MiniMax M3, the OpenAI-compatible API enables thinking when thinking is omitted. Set "thinking": {"type": "disabled"} for a direct response. reasoning_split only changes the response shape; it does not enable or disable thinking.

Anthropic-Compatible Request

curl https://minimax-m2.com/api/v1/messages \
  -H "content-type: application/json" \
  -H "x-api-key: $MINIMAX_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "MiniMax-M3",
    "max_tokens": 2048,
    "thinking": {"type": "adaptive"},
    "messages": [
      {"role": "user", "content": [{"type": "text", "text": "Review this architecture."}]}
    ]
  }'

For MiniMax M3, the Anthropic-compatible API disables thinking when thinking is omitted. Set "thinking": {"type": "adaptive"} to enable thinking blocks. In tool-use conversations, append the complete assistant content array to later turns, including thinking, text, and tool_use blocks.

Multimodal Input

M3 content blocks are preserved by both compatibility endpoints.

• Images: JPEG, PNG, GIF, WEBP; URL or base64 inputs up to 10 MB
• Videos: MP4, AVI, MOV, MKV; URL or base64 inputs up to 50 MB
• Request body: up to 64 MB in the official API
• Larger videos: upload through the official Files API and reference mm_file://{file_id}; official file uploads support videos up to 512 MB

The hosted Chat composer remains text-oriented. Use the API endpoints for image and video requests.

Effective Pricing

This gateway keeps a separate billing identity for M3. Default hosted rates are configured as a flat long-context-safe profile:

These are this deployment's customer billing defaults, not a frozen statement of MiniMax upstream pricing. Query the running deployment before cost planning:

GET /api/v1/models

MiniMax upstream rates vary by input length, cache usage, and service tier. Refer to the live official pricing page for upstream rates.

This gateway currently supports the standard service tier. Requests with service_tier: "priority" return HTTP 400 before an upstream request is made because the billing ledger does not yet record tier-adjusted rates.

Production Validation

1. Confirm MiniMax-M3 appears in GET /api/v1/models.
2. Run a low-token text request through the protocol your application uses.
3. Add representative tool-use and multimodal cases if your product depends on them.
4. Preserve complete assistant content blocks across tool turns.
5. Compare quality, latency, token usage, and failure recovery in your production scaffold before changing defaults.