MiniMaxAI open-sourced MiniMax-M3.1-Flash, a 428B-parameter multimodal MoE model for agentic tasks
Model: MiniMax-M3.1-Flash, MiniMax-M3 · availability, license and releases
MiniMaxAI released MiniMax-M3.1-Flash, a native multimodal mixture-of-experts model with 428B total parameters and 23B activated per token, on Hugging Face. It supports text, image, and video inputs with a context window of up to 1M tokens. The model is available under an 'other' license.
China context
- Original name
- 稀宇科技
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the model weights from Hugging Face and run it locally or on their own infrastructure, subject to the 'other' license terms.
- For investors
- The release of a high-parameter open-weights model by MiniMax may intensify competition in the open-weights segment, potentially affecting pricing and adoption of proprietary APIs.
MiniMaxAI released MiniMax-M3.1-Flash on Hugging Face. The model is a native multimodal mixture-of-experts (MoE) model designed for coding, agentic workflows, computer use, and cybersecurity. It supports text, image, and video inputs with a context window of up to 1M tokens. The architecture includes sparse attention throughout, quantization-aware training, and DSpark speculative decoding. The model has approximately 428B total parameters with 23B activated per token. Reported benchmark scores include 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. The model card includes a comparison table with other models such as DeepSeek V4.1 Flash-1017, DeepSeek V4.1 Flash, Kimi K3, GLM-5.3, GPT-5.6 Sol, and Opus 5.
MiniMax-M3.1-Flash uses a native multimodal mixture-of-experts architecture with sparse attention throughout the model, quantization-aware training, and a DSpark speculative decoding draft head. The model has approximately 428B total parameters and 23B activated per token, supporting a context window of up to 1M tokens. The model card reports benchmark scores including 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. These scores are company-reported and have not been independently verified.
Developers outside China can now access MiniMax-M3.1-Flash weights on Hugging Face, potentially reducing reliance on API-only access for agentic and coding tasks. The model's reported performance on agentic benchmarks may pressure other open-weights providers to release comparable models.
The release of MiniMax-M3.1-Flash provides developers with an open-weights option for multimodal agentic applications, potentially lowering inference costs compared to API-only models. The model's focus on coding, agentic workflows, and cybersecurity may attract enterprise users seeking customizable solutions.
A specific signal to watch is whether independent evaluations confirm the reported benchmark scores, particularly on agentic tasks like OSWorld V2 and CyberGym. Another signal is whether MiniMax releases additional model sizes or updates to the M3.1 family.