MiniMax M3.1 Flash open weights are now downloadable from Hugging Face, giving researchers and infrastructure teams direct access to the 428-billion-parameter model behind the company’s recently launched API. The release fulfills a promise made four days earlier, although the approximately 252 GB artefact remains far beyond a typical workstation or single consumer GPU.
The repository contains 48 safetensors weight shards, tokenizer and configuration files, and a separate DSpark speculative-decoding draft head. There is no approval gate before the download. That makes this a genuine open-weights release, but it is not the same as software distributed under a standard permissive open-source licence.
What the MiniMax M3.1 Flash open weights contain
MiniMax describes M3.1 Flash as a mixture-of-experts model with approximately 428 billion total parameters and 23 billion activated for each token. The smaller active count explains why the model can process tokens efficiently, but it does not reduce the size of the complete download. Developers still need enough aggregate memory, storage and interconnect capacity to hold the weights and run the model.
The shipped files total roughly 252.4 GB. The model card identifies a mixed precision design: routed experts and parts of the sparse-attention key-value path use NVFP4, shared experts use MXFP8, and the query path remains in FP8. This compressed representation is important because unquantized BF16 weights would occupy about 856 GB before runtime buffers, metadata and the key-value cache are considered.
Those estimates are arithmetic rather than a tested minimum deployment specification. Actual requirements will depend on the serving framework, batch size, context length, quantization method, accelerator topology and available memory. The model card recommends SGLang and vLLM and links to deployment recipes, giving experienced operators a starting point rather than a guarantee that existing infrastructure can support the model.
Why the release changes the evaluation process
The most consequential shift is not another benchmark claim. It is the ability to test the exact downloadable build on hardware controlled by somebody other than MiniMax. Scores such as 76.2 on BankerToolBench and 82.2 on CyberGym remain company-reported results, even though the evaluation notes now specify tools, judge models and context settings.
Independent teams can now compare those hosted results with the open build, repeat the evaluations and inspect components such as MiniMax Sparse Attention, quantization-aware training and DSpark speculative decoding. The draft head is included rather than described only in marketing material, allowing researchers to examine how the acceleration mechanism is packaged and whether it works as expected in their own stack.
That work will produce more meaningful evidence than launch coverage alone. It can reveal whether performance depends on particular serving settings, whether efficiency holds under sustained traffic and whether reported agentic gains survive outside MiniMax’s environment. Until those tests appear, the release should be treated as the beginning of an audit rather than final validation of the model.
Open weights are not automatically open source
The repository uses the MiniMax Community License. Its basic grant covers non-commercial use, copying, modification, merging, publication, distribution and sublicensing. Commercial use is permitted, but it is not unrestricted. Products and services must prominently display “Built with MiniMax M3.1-Flash,” users must provide a one-time notice to MiniMax, and products or services generating more than US$20 million in annual revenue require prior written authorization.
The licence also defines commercial use broadly enough to cover paid hosted services, API access and commercial deployment of fine-tuned or modified derivatives. Its prohibited-use appendix includes military purposes and harmful generation of false or misleading information. The licence does not impose a visible territory restriction, but teams operating across multiple jurisdictions still need to review local AI and export rules separately.
For procurement purposes, the distinction is decisive. MIT, Apache-2.0 and similar licences do not attach the same attribution, notice and revenue-authorization conditions. The required attribution wording should also be checked directly against the current licence rather than copied from summaries of an earlier MiniMax release.
Infrastructure support requires verification
MiniMax announced support across a broad hardware and cloud ecosystem, including NVIDIA, AMD and Huawei Ascend chips, several cloud platforms, and serving frameworks such as vLLM, SGLang and FlagOS. The release is still primarily aimed at organisations with substantial compute resources. It is not a desktop model for users expecting an ordinary GPU to run a 428B-parameter system.
Announced compatibility is also weaker than demonstrated availability. Release coverage names production relationships that could not all be confirmed in the providers’ own documentation on release day. The independently verifiable route is self-hosting with the tools directly referenced by the model card. Teams without suitable infrastructure should first compare the hosted API’s price with the operational cost of running the weights themselves.
The hosted rate remains reported at US$0.15 per million input tokens, US$0.60 per million output tokens and US$0.015 per million cached input tokens. For a service with modest traffic, avoiding cluster procurement, monitoring and specialist operations staff may make the API substantially more practical than downloading 252 GB.
Usage figures need a clearer measurement method
Early reporting attached sharply different first-72-hour token totals to the launch. Some coverage cited more than one trillion tokens, while headlines and republications used more than 100 trillion. A total above 100 trillion would technically satisfy a claim above one trillion, but neither version explains which routes, time windows or serving logs were included.
MiniMax also stated that its anonymous Space Bunny build processed more than 200 trillion tokens during a stealth period, while OpenRouter reported 56.7 trillion tokens on its own platform over ten days. These figures may cover different channels and dates, so they need not directly contradict one another. They nevertheless leave readers without a consistent method for comparing demand, throughput or real-world adoption.
The model card also specifies a context window of up to one million input tokens and an output ceiling of 524,288 tokens. Those are substantial limits for coding, document analysis and long-running agents, but actual latency and reliability at extreme context lengths remain questions for independent testing.
Who should download the weights now?
Agent developers, research teams and infrastructure providers that need control over hosting, fine-tuning or data governance now have something concrete to evaluate. They can inspect the exact artefact, build serving environments and compare costs against the hosted endpoint. Their first practical task is capacity planning: estimating storage transfer, multi-accelerator memory requirements, quantization tolerance and the engineering effort needed to keep the system available.
How quickly teams can audit, run and adapt MiniMax M3.1 Flash open weights will decide whether the release changes the open-model landscape or simply adds another large checkpoint to a crowded field.
Consumers hoping to run M3.1 Flash locally should not mistake open weights for a small download. The most important development is the availability of an auditable build, not immediate access to a local model. For now, the strongest conclusion is that MiniMax has delivered the files and enabled outside measurement, while independent benchmarks and real deployment evidence remain the decisive tests ahead for MiniMax M3.1 Flash open weights.