MiniMax M3.1 Flash weights and API limits

Listen to this article

MiniMax released M3.1 Flash on October 19, making the model available through its API, M Plan and MiniMax Code. The company’s announcement puts coding and agent work involving several steps at the center of the release. At launch, downloadable weights and a Pro version were planned for later.

October 23 update brings downloadable weights

MiniMax’s official weight repository now lists a 252 GB release and a weight set labeled as 48 shards. Its initial commit is dated October 23. That is the repository artifact date, not proof of the exact moment it became public. ByteForward inspected the file listings and download links but has not downloaded the weights or run inference.

The model card describes approximately 428 billion total parameters, with 23 billion activated per token. MiniMax says the model accepts text, images and video and produces text, with context up to one million tokens depending on GPU memory and serving configuration. Sparse attention, quantization aware training and an included DSpark draft head are central to its efficiency design.

The community license has commercial conditions

The MiniMax Community License is not MIT. Commercial products and services must prominently display “Built with MiniMax M3.1-Flash”. If those products and services generate more than $20 million in yearly revenue, prior written authorization from MiniMax is required. Otherwise, commercial users must send a one time notice. The terms also apply to derivatives and prohibit uses including military purposes. Read the actual license before deployment rather than treating downloadable weights as unrestricted open source.

Serving limits depend on the hardware

The vLLM recipe updated October 23 calls for version 0.33.0 or later and a mandatory block size of 128. Its prerequisites specify four way tensor parallel serving on NVIDIA Blackwell or AMD MI355X, with ROCm 7.2 or later for AMD, and eight way serving on NVIDIA H100 or H200.

The Hopper example sets context to 262,144 tokens, disables video inputs while retaining images, and does not enable DSpark. The Blackwell and AMD examples use the draft head. These are documented configurations, not performance or minimum memory requirements verified by ByteForward. A one million token headline is not the setting of every deployment.

For local inference, MiniMax recommends temperature 1.0 and top_p 0.95, with reasoning effort set in the chat template. These local settings should not be confused with the hosted Anthropic compatible API behavior described below.

The practical change is control over deployment, with hardware and license obligations attached. The API and hosted access details below retain their October 20 documentation check date. This update does not claim a fresh test of those services.

This is a separate development from the earlier M3.1 Flash Preview promotion covered by ByteForward, which offered subscribers unlimited use inside MiniMax Code from October 1 through October 7.

What the API prices mean

MiniMax’s current pricing table lists $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. All prices are in US dollars. The uncached input and output rates are half the standard M3 rates listed for inputs of up to 512,000 tokens.

Across multiple requests, a workload totaling one million uncached input tokens and 100,000 output tokens costs $0.21 in model usage. If the same input volume qualified entirely for the cached rate, the arithmetic would be $0.075. These are worked examples, not estimates for a completed coding task. Retries and additional calls change the total.

Benchmarks need their labels

MiniMax reports 73.7% on DeepSWE v1.1, 76.2 on BankerToolBench and 61.7 on OSWorld V2. Those are the developer’s published results. ByteForward has not reproduced them, and they do not establish how reliably a particular repository or business workflow will run.

Reasoning settings matter during integration

The Anthropic compatible API documentation lists MiniMax-M3.1-Flash with a one million token context window. Reasoning is always enabled. Trying to disable it returns an error.

The output_config.effort setting accepts low, medium, high, xhigh or max, defaulting to max. Higher settings produce more reasoning tokens and add latency, according to MiniMax.

Reasoning counts toward max_tokens. A small allowance can end the response before answer text appears. Multiple turn tool use also requires preserving the full assistant response, including thinking and tool use blocks.

Temperature, top_p, stop sequences and several other Anthropic parameters are ignored. M3.1 Flash does not support service_tier. Check these differences when migrating an existing integration.

A free trial route has boundaries

OpenCode lists minimax-m3.1-flash-free with free input, output and cached reads for a limited time. It says that route’s provider retains no data and does not use it for model training. Those statements apply to the listed route. They are not a blanket promise about every MiniMax service.

OpenCode’s wider account settings still matter. Its documentation describes automatic $20 reloads when the balance falls below $5, with controls to change or disable reloads. Workspace and member usage limits are available, but an enabled reload can make the amount charged exceed the usage limit. Review these settings before adding paid models to an experiment.

For teams already using Vercel AI Gateway, the model identifier is minimax/minimax-m3.1-flash. Its routing controls distinguish an ordered preference from an exclusive provider list. An order can fall back to unlisted providers, whereas an only list restricts the eligible routes.

A useful comparison should record the provider, reasoning setting, tokens, elapsed time and accepted result for the same tasks. That makes it possible to distinguish a lower token price from a lower cost per successful job. Documentation and prices were checked October 20, 2026. ByteForward did not run the model.