MiniMax M3.1 Flash Weights Are Out: 428B, 252 GB, and a $20M Licence Line

The MiniMax M3.1 Flash weights went up on Hugging Face on 23 October 2026, four days after the model itself launched on 19 October. The download is 428 billion parameters in total with 23 billion active per token, and the repository is 252.4 GB: 48 weight shards plus a 2.3 GB speculative-decoding drafter. It is a real open-weights release, and it is not an open-source one. The licence is MiniMax's own community licence, which asks for attribution on every commercial deployment and for prior written authorisation from MiniMax once annual revenue passes $20 million.

That distinction matters more than the parameter count. Anyone who read "open weights" and assumed MIT or Apache-2.0 has a procurement surprise coming. Anyone who only needs to call the model should probably ignore the download entirely: the hosted API is still $0.15 per million input tokens and $0.60 per million output, and four days of weight files do not change that arithmetic for a team sending a few billion tokens a month.

What is actually in the repository

The model card answers the specification questions MiniMax's launch post left vague. The headline numbers are the ones to keep straight, because several of them differ from the launch-day summaries that circulated without a model card to check them against.

• Parameters — 428 billion total, 23 billion activated per token, a native multimodal mixture-of-experts

• Download — 252.4 GB total: 48 safetensors shards plus a 2.3 GB DSpark drafter for speculative decoding

• Context — 1,000,000 input tokens; up to 524,288 output tokens

• Inputs and outputs — text, image and video in; text out

• Precision — routed experts and the sparse-attention KV path are trained for NVFP4, with shared experts held in MXFP8; the query path stays FP8

• Serving — vLLM, SGLang and TokenSpeed are the recommended paths, with day-0 vLLM containers for NVIDIA and AMD

• Licence — minimax-community, filed on Hugging Face as license: other

• Hosted price — $0.15 input, $0.60 output, $0.015 cached input per million tokens

The 428B/23B split is the same one MiniMax has been quoting since launch. The 252 GB is new information for anyone planning a deployment: it is the actual size of the artefact you would move through a security review, not a theoretical parameter count multiplied by two bytes. The output ceiling is the other number worth noting. MiniMax's API documentation allows 524,288 output tokens, far above the 65,536-token ceiling on Gemini 3.8 Flash and above DeepSeek V4.1 Flash's 384,000. That is a long-generation capability the launch post did not lead with.

The licence is the part to read twice

MiniMax calls this an open-weights release, and by the ordinary meaning of the phrase — downloadable files, no API key — it is. The licence is where the word "open" starts doing less work. The MiniMax Community Licence grants free use, modification and redistribution for non-commercial purposes. Commercial use is allowed, but it comes with two obligations that MIT does not have.

The first is attribution. A commercial deployment has to display "Built with MiniMax M3" prominently on a related website, user interface, blog post, about page or product documentation. The model is M3.1 Flash; the required string still says M3, which is the sort of detail that gets missed until a legal review catches it.

The second is a revenue threshold. At or below $20 million in annual revenue, a company owes MiniMax a one-time notice by email. Above $20 million, it needs prior written authorisation from MiniMax before commercial use. The licence defines commercial use broadly enough to cover offering a product or service that relies on the model, running a commercial API on top of it, and deploying a fine-tuned or post-trained version for commercial purposes. A hosted inference business built on these weights is squarely inside that definition.

That is not a reason to avoid the release. It is a reason to read the release correctly. For a startup under the threshold, the practical cost is an email and a line of attribution. For a larger company, it is a licensing conversation with a Chinese lab, on a timeline MiniMax controls. Compare that with DeepSeek V4.1 Flash, which ships under MIT and needs no notice at all, and the open-weights field is not level even when both checkpoints are downloadable.

252 GB is the easy number; 76 concurrent sequences is the useful one

The interesting serving detail comes from the vLLM day-0 post, which is more specific than MiniMax's own launch material. The NVFP4 checkpoint is 250.1 GB in vLLM's accounting against 451.5 GB for MiniMax-M3 in MXFP8 — a 44.6% reduction. The KV cache falls from 61,440 bytes per token to 34,560, a 43.75% cut, because the sparse-attention indexer and the attention path both run in NVFP4.

The consequence is concurrency at long context, which is the number that decides whether self-hosting is viable. On four NVIDIA B300 GPUs at 256K-token sequences, vLLM reports 76 concurrent sequences for M3.1 Flash where M3 with an FP8 KV cache fits 28 — about 2.7 times more. At the latency-sensitive end of its GB300 AgentX run, the stack sustains 418.9 output tokens per second per user while serving 38.6K total tokens per second per GPU. At the throughput end it reaches 153.8K total tokens per second per GPU at roughly 127 tokens per second per user. On Rubin the same recipe reports 724.9 tokens per second per user, and on AMD MI355X it reports 331.

Those are framework benchmarks on specific hardware, not independent production measurements. They are still the most useful numbers in the release, because they answer the question a parameter count cannot: how many long-context requests fit on the box you can actually rent. A 428B model that fits 76 concurrent 256K sequences on four GPUs is a different procurement object from a 428B model that fits eight.

What self-hosting costs, and when the API still wins

The break-even on open weights is a utilisation question, not a price question. At $0.60 per million output tokens, a billion output tokens a month costs $600 through MiniMax's API. Four B300 GPUs cost far more than that to rent or own before power, networking and the engineer who keeps the serving stack current. A team that can keep those GPUs busy at the throughput end of vLLM's benchmark is buying tokens at a very different effective rate; a team that needs them for two hours a day is subsidising idle silicon.

The workloads that justify the download are the ones where the API price is not the binding constraint. Data that cannot leave a controlled environment. Fine-tuning runs that need weight access, where vLLM reports end-to-end RL with veRL, FSDP2 and rank-32 LoRA on the NVFP4 checkpoint, and a 200-step GRPO run that lifted AIME-2024 accuracy from 36.3% to 67.5% on 16 GB300s. Latency-sensitive agent loops where the network hop and the provider's queue are part of the p95. And batch jobs large enough that a fixed cluster beats a metered bill.

For everyone else, the hosted route remains the sane default, and the release does not change that. The weights are a strategic option, not a cost cut that arrives by default.

What the weights do not fix

Three problems survive the download intact. The first is reliability. Artificial Analysis's AA-Omniscience evaluation put M3.1 Flash at 16 overall with a 93% hallucination rate: when the model did not know an answer, it answered anyway almost every time rather than abstaining. That is a measured behaviour, not a vendor claim, and it is the kind of thing that matters more in an agent that can act than in a chatbot that can be ignored. Open weights let you fine-tune around it; they do not make it go away at the checkpoint.

The second is benchmark provenance. The launch numbers — DeepSWE v1.1 at 73.7%, Agents' Last Exam at 29.6, OSWorld V2 at 61.7, CyberGym at 82.2, BankerToolBench at 76.2 — are MiniMax's own runs on MiniMax's chosen harnesses. None has been independently reproduced. The model card explicitly warns that third-party quantisations may differ, which is a useful admission: the numbers describe the official weights in MiniMax's configuration, not whatever your serving stack produces after you quantise it further.

The third is control. Thinking cannot be disabled: MiniMax's API rejects thinking: {"type": "disabled"} with HTTP 400, and the only lever is the five-rung effort ladder from low to max, with max as the default when the field is omitted. Thinking tokens are billed as output and count against max_tokens. On the Intelligence Index the model consumed roughly 180 million output tokens, well above the median, which is why its $0.22 cost per completed task is better than its per-token rate suggests but still shaped by verbosity.

How to call it now

MiniMax M3.1 Flash is available through MiniMax's own API, M Plan and MiniMax Code, and through several third-party platforms. It is not on the OrcaRouter catalogue today, and this article is not a claim that it is. What OrcaRouter does route is the rest of the family and the models it is usually shortlisted against: MiniMax-M3, DeepSeek V4.1 Flash and Gemini 3.8 Flash all answer on one key, at provider list price with 0% markup, with automatic failover between routes. If you are testing M3.1 Flash against an incumbent, that gives you the incumbent side of the comparison without a second contract, and it gives you somewhere to fail over to while the new model accumulates operational history.

If you are self-hosting, the vLLM container is the shortest path, and the NVFP4 recipe is the one the numbers above describe. If you are calling the API, the release changes little this week except your procurement options.

The honest summary is that the weights are real, the serving story is stronger than expected, and the licence is more restrictive than the phrase "open weights" prepares you for. MiniMax has shipped a genuinely downloadable 428B model with a four-GPU recipe and a 252 GB footprint. It has not shipped this model under MIT, and the difference will matter to exactly the companies big enough to care about the $20 million line.