On October 23, MiniMax announced that it is publishing the model weights of M3.1 Flash, available on Hugging Face and ModelScope at the same time. The company puts the size at roughly 428B total parameters, with 23B active for each token.
The model first appeared in late September as a preview behind the company's API, and at that point there were no corresponding weights on Hugging Face. According to the model card, it is a natively multimodal mixture-of-experts model that takes text, images and video as input and produces text. It is aimed at coding, agentic tasks, computer use and cybersecurity.
What the model card says
There are three architectural changes. First, MiniMax's in-house sparse attention (MSA) now runs through the whole model, replacing the full attention that previously sat in the first three layers. Second, both the routed experts and the attention path went through quantization-aware training. Third, the weights ship with a DSpark draft head that speeds up inference through speculative decoding.
The context window reaches 1 million tokens, though the model card explicitly notes that the length you can actually use depends on GPU memory and the deployment setup. Reasoning strength is set through reasoning_effort, with five levels: low, medium, high, xhigh and max. The recommended sampling parameters are temperature 1.0 and top_p 0.95.
For deployment, the company supports SGLang, vLLM and TokenSpeed, provides images for SGLang and Docker Model Runner, and routes multi-chip setups through FlagOS. The model card does not recommend a specific GPU model or card count.
The license is listed as minimax-community. The actual commercial terms are in the repository's LICENSE file, so teams planning to build a product on the model should read them first.
One more detail in the model card does not add up. The Safetensors metadata read by the Hugging Face page shows about 220B parameters, nearly half short of the 428B the company cites. A rough guess is that this relates to the storage format after quantization, but MiniMax has not commented on it so far.
Who it is compared against
The results the company chose to highlight are concentrated in agentic and security benchmarks:
| Benchmark | M3.1 Flash | Opus 5 as listed in the model card |
|---|---|---|
| BankerToolBench | 76.2 | 76.2 |
| DeepSWE v1.1 | 73.7 | — |
| OSWorld 2.0 | 61.7 | 70.6 |
| Terminal Bench 4.0 | 38.1 | 51.8 |
| CyberGym | 82.2 | — |
On BankerToolBench it ties with Opus 5; on OSWorld 2.0 it trails by about 9 points, and on Terminal Bench 4.0 by 13.7 points. Other published scores include Agents' Last Exam at 29.6, HLE (without tools) at 43.4, BabyVision (with tools) at 91.0 and CharXiv RQ at 82.5. All of these are MiniMax's own results using the official weights. The model card warns that scores for third-party quantized versions may differ, and independent evaluations are not yet available.
Judging by the comparisons the company picked, an open-weight model with 23B active parameters still sits some distance behind closed flagships on long-chain tasks such as terminal and computer operation, while it has caught up in a narrow area like financial tool calling.
From M3 to M3.1 Flash
According to public information, MiniMax-M3 was released on June 1 with the same 428B total and 23B active parameters, and it already put million-token context and sparse attention forward as selling points. M3.1 Flash keeps the parameter count unchanged. The changes are the swapped attention layers, quantization-aware training and the speculative decoding head, with the idea of running a model of the same size more cheaply.
The timeline shows the pace as well: a preview in late September, a Hugging Face repository created on October 17, third-party routing services putting up date-suffixed endpoints around October 19, and the weights made public on October 23. At the time of writing, the repository had been downloaded a little over a thousand times.
For developers in China, the simultaneous ModelScope listing spares them from pulling several hundred GB of weights from overseas. The real barrier to self-hosting is hardware: even a rough 4-bit estimate puts the weights of a 428B-parameter model at over 200 GB of GPU memory, before counting the KV cache for long context, so it takes several 80GB cards to fit. For small and mid-sized teams, calling the API remains the more realistic option.
Sources: MiniMax official announcement, Hugging Face model card, CocoLoop, OpenRouter; used to verify parameter size, context length, license name and model-card benchmark scores.