MiniMax Opens M3.1 Flash; $20M+ Products Need Its Sign-Off

MiniMax published the weights of its M3.1 Flash model on 23 October 2026. “MiniMax M3.1 Flash weights are now public. 428B parameters total, 23B activated per token,” the Chinese AI company wrote on X at 01:06 UTC, linking to the model on Hugging Face and ModelScope. The licence file in the Hugging Face repository grants use free of charge “for non-commercial purposes”; commercial products must display “Built with MiniMax M3.1-Flash”, and those generating more than $20 million in yearly revenue need MiniMax’s prior written authorization.

In plain words, open weights means the trained model file itself can be downloaded and run on your own hardware, instead of only being rented through MiniMax’s paid API. Active parameters are the share of the model that does the work for each word it produces: this model is built from many specialist parts, and the model card says about 23 billion of its roughly 428 billion parameters are used per token. By our arithmetic, that is about 5.4%.

Business Pill · OPEN WEIGHTS

A short Business Engineer explainer on open weights: renting a model versus running it yourself. It teaches the general idea only and says nothing about any company in this piece.

What MiniMax Released

The model card describes MiniMax-M3.1-Flash as “a native multimodal mixture-of-experts model for coding, agentic workflows, computer use, and cybersecurity.” It accepts text, image and video inputs and returns text, with a context window of up to 1,000,000 tokens, according to the card, which adds that usable context length depends on GPU memory and the serving setup.

• Release date. MiniMax had said on 19 October that “Model weights will arrive on Oct 23”; the weights post came on that date.
• Where to run it. The card recommends the SGLang, vLLM and TokenSpeed serving frameworks. Together AI was “a day-0 launch partner,” MiniMax wrote on X at 02:06 UTC.
• What changed from M3. The card lists sparse attention throughout, quantization-aware training and DSpark speculative decoding, which it says improve inference efficiency.

MiniMax’s own benchmark scores for M3.1 Flash, from its X post and Hugging Face model card, 23 October 2026.

What Was Posted on X

MiniMax’s own post at 01:06 UTC on 23 October announces that the M3.1 Flash weights are public, gives the 428B total and 23B active parameter counts and the benchmark scores this piece reports as MiniMax’s own, and links the Hugging Face and ModelScope pages. It quotes MiniMax’s 19 October post, which said Space Bunny was an earlier version of M3.1 Flash and that the weights would arrive on 23 October.

An hour later, MiniMax thanked Together AI for being a day-0 launch partner, writing that Together brings M3.1 Flash to production on its cloud: “Open weights, no GPUs to manage.” Neither post mentions the licence terms set out in the table above.

MiniMax posted on X that M3.1 Flash is now on Tenstorrent Cloud, running on Tensix silicon with Ethernet scale-out and an open-source stack from the team led by Jim Keller.

The Data Point: What the Licence Allows

MiniMax’s posts call the weights open. The licence file attached to them sets the terms, and Hugging Face’s listing tags the model with an “other” licence named minimax-community, not a standard open-source licence.

Source: the LICENSE file in MiniMax’s Hugging Face repository for MiniMax-M3.1-Flash, read 23 October 2026. Commercial use, as the file defines it, includes offering paid products or services that rely on the model, commercial use of its APIs, and deploying fine-tuned or otherwise modified versions for any commercial purpose.

The licence file we read on 23 October does not list excluded countries or regions.

The Benchmarks, as MiniMax Reports Them

MiniMax’s X post lists “76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, and 82.2 on CyberGym.” Its model card sets those scores beside rival models’ in one table. These are MiniMax’s own results.

Source: the benchmark table in MiniMax’s Hugging Face model card, 23 October 2026. All scores, including the rivals’, are as MiniMax reports them, under evaluation settings MiniMax lists; a dash means no score in MiniMax’s table. We have not seen independent results.

In MiniMax’s table, M3.1 Flash matches Opus 5 on BankerToolBench at 76.2 and trails it on DeepSWE v1.1, OSWorld 2.0 and Agents’ Last Exam; the table gives no CyberGym score for Opus 5.

Where Space Bunny Fits

MiniMax said on 19 October that Space Bunny, a model it says “has topped global leaderboards with 200T+ tokens,” “was an earlier version of M3.1 Flash.” Its 23 October weights post quotes that earlier announcement, which priced the API at a flat $0.15 per million input tokens and $0.60 per million output tokens.

Connecting the Coverage

Our M3.1 Flash pricing piece reported on 19 October 2026 that MiniMax listed the model at a flat $0.15 per million input tokens and $0.60 per million output tokens, the same as the off-peak rate DeepSeek lists for DeepSeek V4.1 Flash, and that MiniMax said the weights would arrive on 23 October. What’s new: the weights are now public, with the licence terms above.

Our export-filing piece reported that Chinese state media said on 22 October 2026 that DeepSeek and MiniMax had completed registration under China’s new pre-export filing rule for AI technology and had completed open-sourcing and overseas licensing. MiniMax’s weights post came the next day, at 01:06 UTC on 23 October, and does not mention the registration.

Our DeepSeek and Huawei funding piece reported, citing Bloomberg on 22 October 2026, that DeepSeek raised about 100 billion yuan ($14.98 billion) in an equity round. MiniMax’s model card benchmarks M3.1 Flash against DeepSeek V4.1 Flash.

The Structural Read

As we read it, MiniMax now offers the same model two ways: as a metered API at the flat $0.15 and $0.60 per million tokens it announced on 19 October, and as downloadable weights governed by its own licence.

As we read it, the $20 million line sorts users by size: smaller commercial products need a notice and the “Built with” label, while larger ones need MiniMax’s written authorization, according to the licence file.

As we read it, the benchmark case MiniMax makes is about agent work at a lower price; the scores it reports, including the rivals’, are its own.

Three Implications

FOR DEVELOPERS AND RESEARCHERS The licence grants use free of charge for non-commercial purposes, with the copyright and permission notice kept in all copies, according to the LICENSE file.

FOR COMPANIES BUILDING PRODUCTS Commercial use requires displaying “Built with MiniMax M3.1-Flash” and either a one-time notice to MiniMax or, above $20 million in yearly revenue, its prior written authorization, per the licence file.

FOR ANYONE FINE-TUNING IT The licence counts deploying post-trained, fine-tuned or otherwise modified versions for any commercial purpose as commercial use.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework’s core pattern: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.

As we read it, MiniMax opens the weights and keeps a narrower layer under its control: the right to authorize commercial use above $20 million in yearly revenue, per its licence file. The framework was published in March 2026 and does not discuss MiniMax.

What Is Not Established

The benchmark scores are MiniMax’s own, including the scores it lists for rival models; we have not seen independent results. We read the licence file and the model card as published on Hugging Face on 23 October 2026. We did not contact MiniMax. Nothing here is investment advice.

The Bottom Line

MiniMax published the M3.1 Flash weights on 23 October 2026: about 428 billion parameters, 23 billion active per token, with a context window of up to 1 million tokens. The licence grants free use for non-commercial purposes; commercial products must say they are built with the model, and those above $20 million in yearly revenue need MiniMax’s prior written authorization.