MiniMax M3.1 Flash is a released multimodal mixture-of-experts model for coding and tool-driven workflows. It accepts text, images and video and produces text. The important November update is that official weights are downloadable, so older descriptions saying the weights will arrive later are obsolete.

M3.1 Flash has its own checkpoint, inference recipes and license consequences. The original M3 guide describes another release and should not be silently relabeled. This page focuses on choosing a route, planning deployment and understanding the limits of the available evidence. We have not independently tested the checkpoint or reproduced MiniMax’s benchmarks.

Identity, modalities and context

Property	Official model-card specification
Checkpoint	MiniMaxAI/MiniMax-M3.1-Flash
Architecture	Native multimodal mixture of experts
Parameters	Approximately 428B total, 23B active per token
Inputs	Text, image, video
Output	Text
Context	Up to 1,000,000 tokens
Attention	MiniMax Sparse Attention throughout
Draft head	DSpark speculative decoding
License	MiniMax M3.1 Flash community license

Video input means the model can analyze supplied video. It does not mean this is a video-generation endpoint. Likewise, computer-use training does not supply a browser or desktop executor. Your application chooses which observations to provide and which proposed actions may execute.

The total parameter count matters for deployment even though only a subset is active for each token. Do not size memory as if only 23B parameters must be stored. Caches, precision, vision processing and concurrency change the actual serving requirement. Our GPU memory planning guide gives the surrounding framework.

What changed from M3

MiniMax describes sparse attention throughout the model, quantization-aware training and the included DSpark draft head as important efficiency changes. These are architectural and training descriptions. Their effect depends on the runtime implementation and workload, rather than being a universal speed multiplier.

The model card reports results for coding, tool use, computer interaction and cybersecurity evaluations. Those scores belong to MiniMax’s published evaluation setup. They are not AI Made Tools measurements and should not be merged with unrelated benchmark versions to manufacture a ranking.

For migration, preserve the exact M3 baseline, client settings and acceptance tests. Run the new model with the same tools and task budget. Record successful outcomes, review effort and wall-clock time before deciding whether Flash improves the job you actually run. Our M3 coding guide remains useful context for that comparison.

Hosted API and client integration

Use the official text-generation documentation to configure the direct MiniMax route. A MiniMax subscription, direct API key and third-party router account are separate access arrangements. Verify the exact model identifier and supported request format at the chosen endpoint.

For production planning, check the current official price page and billable cache categories rather than transplanting M3 prices or assuming a subscription includes unrestricted API traffic. This guide does not present a separately verified current direct price. Get API Pulse maintains the structured provider records; confirm the route and verification date before using a figure in a budget.

A model returning a tool call does not mean the tool ran. Your executor must validate arguments, check permissions and report the actual result. Separate failed tools from model failures in logs. The agent security guide and regression testing guide cover these controls.

Downloading and serving the released weights

The model card documents downloading with the Hugging Face CLI:

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

This is a documented download command, not a claim that we executed inference or that the model fits your machine. Review available storage and download size before running it. For reproducibility, pin the chosen repository revision when building a real deployment.

MiniMax links SGLang and vLLM recipes, plus a FlagOS multi-chip path. Those recipes are more useful than an unqualified generic launch command because they encode architecture and hardware assumptions. A runtime that supports M3 does not automatically support all M3.1 features.

Start with a short request, verify the response and then enable tools and multimodal inputs separately. Test the exact chat template and parser. Only after that should you raise context and concurrency. A million-token specification does not promise that full-window serving is economical on your configuration.

Reasoning and inference settings

The model card recommends temperature 1.0 and top-p 0.95. It lists low, medium, high, xhigh and max reasoning effort in the chat template. These are vendor recommendations, not proof that every hosted endpoint accepts identical parameters.

Set a completion budget and inspect actual usage. More reasoning can change latency and output consumption. An unsupported client parameter may be rejected or ignored, so validate what the selected service sends and returns. Keep a known-good configuration available for rollback.

The license changes the deployment decision

Read the actual M3.1 Flash license. It contains commercial attribution and notification requirements, including displaying “Built with MiniMax M3.1-Flash.” Products and services above the stated $20 million yearly-revenue threshold require prior written authorization; the license describes a one-time notice for other commercial use.

That is materially different from assuming an unrestricted MIT or Apache license. Check the complete definitions for your application and retain a copy of the terms attached to the deployed revision. Runtime packages and service contracts do not replace the model’s own license conditions.

My take

M3.1 Flash has enough substance for a distinct evaluation: released multimodal weights, documented serving options and a license that affects deployment. Start hosted if you need a capability baseline. Consider self-hosting when control or economics justifies operating a large MoE, after accounting for hardware, support and license conditions. The name Flash does not itself establish low total cost.

FAQ

Are the weights available?

Yes. The official MiniMaxAI repository hosts M3.1 Flash. Earlier promised-release wording is no longer the current status.

Does it generate video?

The documented checkpoint accepts video input and produces text. Do not confuse video analysis with video generation.

Is it a 23B model for memory planning?

No. Approximately 23B parameters are active per token, while the model has approximately 428B total. Plan the complete runtime footprint.

Can I reuse the old M3 license assumptions?

No. Review the M3.1 Flash license, including its attribution, notice and authorization conditions.