The promise, kept on the reported schedule

When M3.1 Flash launched on October 19, MiniMax's blog said the weights “will be released later” and gave no date. Reports at the time pointed to October 23. MiniMax has released the M3.1 Flash weights on the date earlier third-party reports had anticipated — and unusually for a launch week, the central fact does not rest on the announcement coverage. We opened the repository. MiniMaxAI/MiniMax-M3.1-Flash on Hugging Face contains the complete model: 48 safetensors weight shards, the DSpark draft head, the tokenizer and configuration files, about 252 GB in total, and no approval gate in front of the download.

That sequence matters because it is the verifiable part of the launch. The benchmark table has not changed since Monday and remains MiniMax-reported; the model card repeats the launch figures, including 76.2 on BankerToolBench and 82.2 on CyberGym, and its evaluation notes name the setups — BankerToolBench run in OpenCode 1.17.18 with a MiniMax-M2.7 judge, CyberGym in Claude Code 2.1.263 at a 1M context — and state that the results are for the official weights. What changed today is not a claim. It is an artefact anyone can download, inspect and benchmark on hardware MiniMax does not control. Whatever independent evaluators find in the coming days will be the first numbers in this model's story that its maker did not produce.

A Flash model you cannot put on a laptop

The name invites a misunderstanding worth correcting plainly. M3.1 Flash is a mixture-of-experts model with approximately 428 billion parameters in total and about 23 billion activated per token. The active count is what makes it fast and cheap to serve. It does not make the download small.

At 428 billion total parameters, the full model is far beyond the practical capacity of a typical single-GPU consumer workstation. Self-hosting will depend on the available weight formats, quantisation and memory configuration; the release reports do not provide a verified minimum hardware specification. The model card does pin down the shipped format: the download totals about 252 GB because the routed experts are quantized during training — the configuration specifies MXFP8 quantization with an NVFP4 format for the MoE path. The unquantized arithmetic shows why the quantization is not optional at this scale: 428 billion parameters at BF16 would need roughly 856 GB for the weights alone, and a generic 4-bit representation about 214 GB. Those figures are arithmetic, not measured deployment requirements, and they exclude quantisation metadata, runtime buffers and the KV cache.

Read the day-one support list with that in mind. The announcement, as reported by Chaincatcher, names ten chip makers, from Huawei Ascend to NVIDIA and AMD, six cloud platforms, and serving support from vLLM, SGLang and FlagOS, with distribution via Hugging Face and ModelScope. The announced chip, cloud and serving-framework support points to infrastructure operators and teams with substantial compute resources as an important target audience. But announced support is not verified deployment. Release coverage names six cloud platforms as part of the initial support ecosystem; the reports do not independently establish that all six had production-ready M3.1 Flash endpoints available on release day. RuntimeWire reports Together AI as the production inference partner paired with the release, citing MiniMax as its source; Together AI's own model documentation, checked on 23 October, lists MiniMax M3 and no M3.1 Flash endpoint. The deployment route the model card itself confirms is self-serving: it recommends SGLang and vLLM and links to a cookbook and a recipe for serving this model with each. Everything beyond that is announced compatibility whose test status is not established.

One launch, two token thresholds

The announcement also carried a usage figure, and the coverage cannot keep it straight. The same release coverage uses two markedly different thresholds for first-72-hour token usage: more than one trillion in Chaincatcher's body text, and more than 100 trillion in its headline and in the KuCoin and Cointime republications. The higher figure has not been independently established, and the reports do not explain the difference. Until MiniMax supplies a clearly scoped figure, neither threshold should be treated as a verified usage total.

Precision matters here. Both figures are lower-bound claims: a total above 100 trillion would also be above one trillion, so the body text does not disprove the headline. What the coverage offers is a hundredfold difference in claimed scale for the same window, with no stated scope in any version — which routes were counted, whether the stealth Space Bunny period is included, or whose serving logs the total comes from. Context makes the gap harder, not easier, to wave away. MiniMax said the anonymous Space Bunny build processed more than 200 trillion tokens over its stealth run, while OpenRouter's own summary put that build at 56.7 trillion tokens on its platform in ten days; those two are not necessarily contradictory, since one may cover more routes or a different window than the other. A reader now has three totals attached to closely related claims, spanning two orders of magnitude, none with a stated counting method.

What the licence actually says

Two documents decide what this release is worth, and neither is the announcement. The first is the licence in the repository, and it can now be read rather than guessed at. The weights ship under the MiniMax Community License, the licence document in the repository's LICENSE file. Its opening grant is framed for non-commercial purposes — to use, copy, modify, merge, publish, distribute and sublicense the software. Commercial use is not prohibited outright, but it is conditioned: a product or service using the model commercially must prominently display “Built with MiniMax M3.1-Flash”; the user must send MiniMax a one-time notice; and if the products or services concerned generate more than US$20 million in yearly revenue, prior written authorization from MiniMax is required. The licence defines commercial use broadly, expressly including paid hosted and cloud offerings of the model or its APIs, and the commercial deployment of fine-tuned or otherwise modified derivatives. An appendix lists prohibited uses, including military purposes and generating false or misleading information with the intent to cause harm. Just as notable is what this document does not contain: no territory restriction appears in it.

MiniMax has previously used model-specific community licences. The applicable terms for M3.1 Flash must be established from this release's own licence document; restrictions described in coverage of the separate H3 release cannot automatically be applied here. On this licence's own text, “open weights” is accurate and “open source” is not. The word in Chaincatcher's headline, carried into the Cointime and KuCoin republications, is those outlets' terminology, not a description of the document MiniMax actually shipped. For a commercial deployment, the download is the beginning of the process, not the whole of it.

The second document does not exist yet: an independent benchmark of the downloadable build, run on a third-party stack. The hosted model's scores have already drawn outside measurement since Monday. The weights are the version researchers can audit properly, including the efficiency techniques the card names: MiniMax Sparse Attention throughout the model, quantization-aware training, and DSpark speculative decoding, whose draft head ships in the repository as a separate file. That audit is the real event this release sets up. It starts now, and its results will say more about M3.1 Flash than the announcement does.

Sources

  1. [1] MiniMaxAI/MiniMax-M3.1-Flash on Hugging Face — the model card, file listing and LICENSE file, read directly on 23 October 2026: complete weights in 48 shards (about 252 GB) plus the DSpark draft head; MiniMax Community License text; approximately 428B total and 23B activated parameters; sparse attention, quantization-aware training and DSpark; BankerToolBench 76.2 and CyberGym 82.2 with evaluation notes; SGLang and vLLM serving recommendationsRead source
  2. [2] Chaincatcher, “MiniMax M3.1 Flash open source 72 hours processing Token over 100 trillion” (23 October 2026, 01:06 UTC) — the weights-release announcement report: the announced chip, cloud and framework support list, and body text stating first-72-hour tokens “exceeded one trillion” under a headline stating 100 trillion. “Open source” in the headline is the outlet's terminologyRead source
  3. [3] RuntimeWire, “MiniMax opens M3.1 Flash weights with Together AI handling inference” (23 October 2026, 02:43 UTC; primary source cited: MiniMax) — Together AI named as the production inference partner paired with the weights releaseRead source
  4. [4] KuCoin flash news, “The MiniMax M3.1 Flash Model processes over 100 trillion tokens in 72 hours” — republication of the Chaincatcher report carrying the 100-trillion figure and the same support listRead source
  5. [5] Cointime, “MiniMax M3.1 Flash Open Source Processes Over 100 Trillion Tokens in 72 Hours” — second republication carrying the 100-trillion figureRead source
  6. [6] Together AI documentation, “Available models” (checked 23 October 2026) — lists MiniMax M3 among its chat models; no M3.1 Flash endpoint listedRead source
  7. [7] MiniMax, “MiniMax M3.1 Flash: Redefining Agent Performance and Cost” (official blog, 19 October 2026) — the launch this release follows: flat pricing of $0.15 / $0.60 / $0.015 per million tokens and the statement that weights “will be released later”, with no date givenRead source