One-line summary
MiniMax M3.1 Flash (officially open-sourced on October 23, 2026, with 428B total and 23B active parameters) can be used through the Open Platform API, or self-hosted by downloading the weights; for local deployment, use an inference framework such as vLLM that supports MoE sparse activation, and size hardware by active parameters and precision (BF16/MXFP8).
When to use this guide
- Developers and enterprises that need to run MiniMax M3.1 Flash in a private environment (data stays on the intranet, custom inference policies).
- Teams that need lower call cost or higher concurrency: MiniMax states M3.1 Flash delivers nearly 3x faster inference and 70% lower usage cost than the previous generation (release statement of October 19, 2026).
- Individuals who want fast prototyping without buying compute: call the Open Platform API directly.
Steps
Option 1: Open Platform API (fastest)
- Register an account on the MiniMax Open Platform (platform.minimaxi.com).
- Create an API Key in the console, and confirm the M3.1 Flash model ID and current pricing in the docs. At launch, the platform price was CNY 1 per million input tokens and CNY 4 per million output tokens (per Tencent Cloud Developer Community, October 19, 2026; check the platform for the latest).
- Call the chat completion endpoint with model and messages parameters, authenticating with the API Key (Bearer scheme).
Option 2: Self-host with vLLM (example)
- Check hardware: M3.1 Flash has 23B active and 428B total parameters; BF16 weights are on the order of ~800GB, MXFP8 quantization roughly halves that; single-machine serving needs multi-GPU with large VRAM or a large unified-memory setup, following the memory guidance in the official M3 self-hosting doc as a reference scale.
- Get the weights: download the M3.1 Flash repository from the Hugging Face MiniMaxAI org (the open-source announcement was published October 23, 2026; previous-gen M3 weights live at MiniMaxAI/MiniMax-M3-MXFP8, M3.1 Flash per the official repo), and check the repository license (M3 uses the MiniMax Community License; M3.1 Flash per the official repo).
- Install an inference framework: use vLLM (supports MoE sparse activation) or SGLang; example with vLLM:
pip install vllmpython -m vllm.entrypoints.openai.api_server --model <local_weight_path> --tensor-parallel-size <num_gpus>
- After the server starts, connect your application through the OpenAI-compatible endpoint (the local inference service listens on port 8000 by default, with the /v1/chat/completions path), or use vLLM's serving tools for load testing.
- Verify: send a sample Chinese/English request, confirm normal output, and check throughput (tokens/s) and memory usage.
Option 3: Cluster inference (large scale)
- For higher concurrency or larger contexts, deploy vLLM across multiple nodes with pipeline/tensor parallelism; use high-speed RDMA networking; final topology and tuning should follow the vLLM official docs and MiniMax deployment guidance.
Notes
- License: the commercial license for M3.1 Flash follows the actual license in the open-source repository; previous-gen M3 uses the MiniMax Community License, so read the terms (e.g., monthly active user limits) before commercial use.
- Weight version: MiniMax first said weights would be "released later" and open-sourced them on October 23, 2026; confirm the repository revision when downloading to avoid stale weights.
- Data compliance: local deployment keeps data on your machine, but disable unnecessary telemetry and follow local data protection regulations.
- Price verification: platform pricing and model IDs may change; confirm the current docs on platform.minimaxi.com before integrating.
References
- MiniMax Open Platform Docs: Self-hosting MiniMax-M3 link
- Cailianshe (第一财经): MiniMax M3.1 Flash processed over 100T tokens in three days; officially open-sourced today (Oct 23, 2026) link
- Tencent Cloud Developer Community: MiniMax M3.1 Flash launch - CNY 1 in / CNY 4 out per million tokens (Oct 19, 2026) link
- vLLM official docs: Serving link