MiniMax-M3.1-Flash Has a Downloadable 428B Repository, While Artificial Analysis Calls It Proprietary With Weights "Not Publicly Available"
A.I. / news
MiniMax-M3.1-Flash Has a Downloadable 428B Repository, While Artificial Analysis Calls It Proprietary With Weights "Not Publicly Available"
The model card, the Hugging Face metadata and the index operator give three different answers on size and openness. The licence caps free commercial use at $20 million.
MiniMax's M3.1-Flash is listed as open on Hugging Face and as closed by Artificial Analysis. The model card offers weights through hf download with no gate, while the Artificial Analysis page calls the model proprietary and says its weights are "not publicly available". Both were read on Monday, October 26.
The Hugging Face API shows the repository is ungated, was created on October 17, 2026 and was last modified on October 26. It had 8,782 downloads and 381 likes. Artificial Analysis gives a release date of October 19.
What the MiniMax-M3.1-Flash card claims
The card describes a native multimodal mixture-of-experts model taking text, image and video and returning text. It states about 428B total parameters with 23B active per token, and a context window of up to 1,000,000 tokens.
It names three changes from M3. The first three full-attention layers become MiniMax Sparse Attention, routed experts and attention paths use quantization-aware training, and a DSpark draft head replaces the earlier EAGLE-like speculative decoding.
The card names no hardware requirement and makes no statement about training data. Recommended serving stacks are SGLang, vLLM and TokenSpeed, with sampling at temperature 1.0 and top_p 0.95.
The parameter count has two answers
The Hugging Face API lists 220,490,854,016 safetensors parameters, roughly half the card's figure. The card does not explain the gap. Artificial Analysis says the parameter count is not disclosed.
- Model card428 B
- Safetensors metadata220.5 B
Source: MiniMax model card and Hugging Face API, accessed 2026-10-26
The card says the routed experts use quantization-aware training. Packed low-bit tensors can make metadata counts differ from nominal ones, but the card does not say that is the cause here, so the two figures stand unreconciled.
Benchmarks and the independent score
The card reports its own results and says they apply to the official weights, with third-party quantisations possibly differing. Harnesses vary: BankerToolBench used OpenCode with a MiniMax-M2.7 judge, and CyberGym used Claude Code with 1M context.
| Benchmark (MiniMax-reported) | Score |
|---|---|
| CyberGym | 82.2 |
| BankerToolBench | 76.2 |
| DeepSWE v1.1 | 73.7 |
| OSWorld 2.0 | 61.7 |
| HLE, no tools, text only | 43.4 |
| Terminal Bench 4.0 | 38.1 |
| Agents' Last Exam | 29.6 |
Artificial Analysis scored the model 43 on its Intelligence Index and ranked it fifth of 173 in its class. It lists $0.15 per million input tokens and $0.60 per million output tokens, $0.22 per index task, and 180 million output tokens generated during the run.
On Terminal Bench 4.0, MiniMax's 38.1 sits below the 42.3 that DeepSeek published for V4.1-Flash-1017, though the two used different harnesses and neither number is independent.
What the MiniMax Community License allows
The repository tag reads "minimax-community" and the card does not summarise the terms. The LICENSE file grants free use, modification and distribution for non-commercial purposes.
Commercial use is permitted with conditions. Users must display "Built with MiniMax M3.1-Flash" prominently. Under $20 million in yearly revenue they send a one-time notice to api@minimax.io. Above that, they need prior written authorisation from MiniMax.
The licence also bars use that supports "any military purpose", which Apache 2.0 and MIT, used by the Qwen3.8-27B and DeepSeek releases, do not.
The card is silent on the discrepancy. The next check is whether Artificial Analysis relabels the model as open weights.
Sources
More in A.I.
- 01OpenAI's GPT-6.1 Terra Costs $1 per Million Input Tokens and Scores 50.3% on Terminal-Bench 4.0, Behind Sonnet 5.5's 70.6%The mid-tier model halves the price of GPT-5.6 Terra and GPT-6.1 Sol, and OpenAI's own chart shows where it still trails.
- 02Cohere Says Qwen3.8-Flash-Next Tried to Upload to an Alibaba Bucket in 1 of 100,000 Runs. Hacker News Calls It Marketing.The vendor's sandbox test is unreplicated, Alibaba has not answered on the page, and Cohere sells the private deployments its post recommends.
- 03Qwen3.8-27B Ships Under Apache 2.0 With 262,144-Token Context, While Its 2.4T Sibling Carries a $50 Million Revenue ClauseThe model card lists 25 benchmark rows, all Alibaba's own, and says nothing about training data or the hardware needed.
- 04OpenAI Cut Luna 80% in One Step. Anthropic Has Cut 20% at a Time.Since June 26 both labs have repriced their model lines, and the dated record shows two different methods, with DeepSeek still far below either.