Deployment

Download MiniMax M3.1 Flash Weights

Official weight files are available; this guide links the source, explains the license, and sets realistic deployment expectations.

Reviewed Not independently reproduced by us

MiniMax has published the M3.1 Flash checkpoint on its official Hugging Face repository. The repository is the source of truth for files, revisions and license. This page was checked on 23 October 2026; availability and framework support can change.

Download from the official repository

Install the Hugging Face CLI, authenticate if the repository requires it, then download the checkpoint:

hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-Flash

This downloads files; it does not start inference. The repository contains sharded model weights, an index and a dspark/ draft component. Follow the current model card for the exact file set and checksums. Avoid third-party mirrors when you need provenance.

Choose a serving framework

The official vLLM recipe documents tensor parallelism on supported accelerator systems: TP4 on Blackwell or AMD MI355X, and TP8 on NVIDIA Hopper. The SGLang recipe provides a separate supported path. Check each recipe’s current framework version, accelerator and launch flags before provisioning hardware.

The model card describes roughly 428B total parameters and 23B active parameters. Active parameters per token do not describe the checkpoint’s storage requirement; do not estimate disk or accelerator memory from 23B alone. We have not run either recipe ourselves, so this page makes no promise about a particular workstation or card.

Read the license before commercial use

The repository specifies the MiniMax Community License. It permits use under stated conditions and includes notice/display requirements and a prior written authorization threshold for products exceeding $20 million in annual revenue. Read the complete license text for your use case; this summary is not legal advice or a grant of rights.

When hosted access is a better fit

If you do not operate a supported multi-accelerator server, use a hosted route instead: Ollama cloud or the MiniMax API. Downloading open weights removes a provider dependency, but does not remove the hardware, serving or license requirements.

Next: current MiniMax model identifiers or compare M3.1 Flash with DeepSeek V4.1 Flash.

Sources and version notes

Checked October 23, 2026. We label official claims, third-party measurements, our own testing and community reports separately — see the sources policy for what each label means.

Spotted something out of date? Suggest a correction →

Your next step

OllamaOllama offers a cloud route; MiniMax also publishes weights, with deployment requirements covered separately.API referenceTwo compatible endpoints, one model ID, an effort dial that defaults to max, and a 1M-token context window.Model listOne table of every live model, the retirement timeline for the M2.x family, and which one to point your agent at.vs DeepSeek V4.1 FlashSame measurement run, same providers, two very different trade-offs — and one place where the vendor-supplied data caveat matters most.