MiniMax's M3.1 Flash weights bring image and video inputs to a self-hosted checkpoint

MiniMax published M3.1 Flash weights with image and video input support.

MiniMax's public model repository records the initial M3.1 Flash checkpoint on October 23. That revision includes model weight shards, giving developers a downloadable alternative to hosted inference. MiniMax identifies the checkpoint as a mixture-of-experts model for workloads including coding, tool use and computer interaction. Its documentation lists text, image and video inputs, with a context window of up to one million tokens. [Model card](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash) · [Repository history](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/commits/main) · [Initial revision file inventory](https://huggingface.co/api/models/MiniMaxAI/MiniMax-M3.1-Flash/tree/54e4e69)

What downloading the model changes

With the weights on its own infrastructure, a team can choose the runtime, control the model revision and measure results under a repeatable serving configuration. The card includes vLLM and SGLang examples and identifies the license as MiniMax Community. Organizations assessing redistribution or commercial use should read the terms for their intended use. The license is part of the deployment decision alongside the available hardware. [Model documentation and license](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash)

The company reports about 428 billion total parameters, of which about 23 billion are active for each token. Active parameters are an indication of how much of the expert network runs for that token. The full checkpoint still has to be made available to the runtime. A team sizing its servers would also account for image and video processing, the amount of history kept per request and expected concurrency. [Technical specifications](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash)

Evaluating multimodal work

A video test can ask the model to identify where a sequence of events changes and compare its answer against annotated timestamps. A coding evaluation might give it a screenshot of a failure and a repository, then check whether the resulting patch passes tests. Both use images or video, but the second also tests whether the surrounding agent grants and records the right actions. Model output quality and tool authorization are different parts of the system.

The checkpoint makes both kinds of evaluation possible on infrastructure a team controls. The practical comparison with hosted inference is the cost and quality observed on the team's actual media, prompt lengths and workload.