Friday, Nov 13 | --:--
Times of AI

Regions

China Europe India Japan Taiwan United Kingdom United States

Platforms

Alibaba Qwen Amazon Anthropic DeepMind Google Groq Meta Microsoft MiniMax Mistral Moonshot AI NVIDIA OpenAI PrismML Replit Thinking Machines xAI

Sections

Business Engineering Enterprise Frontier Labs Hardware Industry Infrastructure Models Policy Products Research Safety Security Work All articles → Subscribe by email
Back to home
Models Models China

Update: MiniMax Publishes M3.1 Flash Open Weights Under a Community Licence That Limits Commercial Use; AA Scores It About 42

MiniMax published M3.1 Flash's weights on Hugging Face on Oct. 23 under a MiniMax community licence. Commercial users must display attribution, and products above $20M in yearly revenue need MiniMax's written authorization, so the release is open weights, not open source. Artificial Analysis had not yet updated its listing and scores the model about 42. Released Oct. 19 at $0.15/$0.60 per million tokens.

Times of AI Desk 3 min read Shanghai View as Markdown
Engraved flash bolt over a rabbit silhouette and a locked weight plate, violet intaglio.

Updated October 23, 2026: MiniMax has published M3.1 Flash's weights on Hugging Face under its own community licence. That makes it an open-weights model, not open source. Artificial Analysis had not yet updated its listing when we checked. The new section directly below covers this; earlier text follows unchanged.

October 23: weights are out under a MiniMax community licence#

The promised weights arrived on schedule, but under terms that limit commercial use. Free use is restricted to non-commercial purposes. The MiniMaxAI/MiniMax-M3.1-Flash repository on Hugging Face is public and not gated, and the Hub's metadata showed it last modified at 02:26 UTC on Friday, October 23. The model card describes a native multimodal mixture-of-experts model with approximately 428 billion total parameters and 23 billion activated per token, text, image and video inputs, and a context window of up to 1 million tokens. It says the release includes a DSpark speculative-decoding draft head and recommends SGLang, vLLM and TokenSpeed for serving.

The repository's MiniMax Community License grants free use for non-commercial purposes. Commercial users must prominently display "Built with MiniMax M3.1-Flash." Products or services with more than $20 million in yearly revenue need MiniMax's prior written authorization, and smaller ones must send MiniMax a one-time notice. An appendix bars uses including any military purpose. Because commercial use is conditioned on attribution and, above that threshold, on permission, we describe the release as open weights rather than open source.

Yicai reported on October 23 that MiniMax had open-sourced the weights, and that MiniMax said the model processed more than 100 trillion tokens in its first 72 hours online (our translation from Chinese). That figure is MiniMax's claim, and we have not verified it. Yicai also reports that MiniMax said it is working with chipmakers including Huawei Ascend, T-Head, Kunlunxin, Moore Threads, Nvidia and AMD on day-one support.

Limits of the October 23 update#

  • Artificial Analysis had not yet updated its listing when we checked at 03:11 UTC on October 23. Its model page still marked M3.1 Flash as proprietary, with its open-weights flag set to false. The score of about 42 above is unchanged.
  • The model card's approximately 428 billion total parameters does not match the 220 billion that Hugging Face's safetensors metadata reports for the repository. We have not reconciled the two figures.
  • The card lists 82.2 on CyberGym, where MiniMax's launch blog, as reported above, gave 82%. All benchmarks on the card are MiniMax's own except one it says it took from Artificial Analysis.
  • The translation of Yicai's report is ours. We have not tested the weights or checked the licence against any other text.

Another cheap Chinese agent-focused model landed two days after DeepSeek's Flash update — and this one is not open weights yet. MiniMax on Monday, October 19 released M3.1 Flash, which it says delivers nearly 3x faster online inference and 70% lower agentic usage cost than its predecessor. List price is $0.15 per million input tokens, $0.60 per million output and $0.015 cached input, served at 200+ tokens per second (P50 170+), MiniMax says. An early version ran anonymously as "Space Bunny" and processed more than 200 trillion tokens, according to the company.

Independent score#

Artificial Analysis scores M3.1 Flash about 42 on its Intelligence Index (42.71 in the payload we parsed), up from 29 for MiniMax M3, VentureBeat reports, with cost per task down from about $0.51 to $0.22. AA's isOpenWeights flag is false. On AA-Omniscience, VentureBeat says the model was 56% correct with a 93% hallucination rate under AA's definition, for an Omniscience score of 16.

MiniMax's own benches#

MiniMax reports 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 and 82% on CyberGym. Those figures are the company's. The model is available now through the MiniMax API, M Plan and MiniMax Code. "Model weights will be released later," the blog says; VentureBeat reports MiniMax plans to publish them on October 23, with licence terms not yet given.

Limits#

  • Every benchmark except Artificial Analysis is MiniMax self-reported. Weights and licence for October 23 are promises, not a release.

Sources#

Related Coverage

Scroll to continue reading