MiniMax releases MiniMax-M3.1-Flash, a native multimodal MoE with 1M-token context
MiniMax released MiniMax-M3.1-Flash, a native multimodal mixture-of-experts model for coding, agentic workflows, computer use, and cybersecurity. It is available via the MiniMax API and on Hugging Face under the minimax-community license.
Key facts
- Approximately 428B total parameters with 23B activated per token; native multimodal MoE architecture.
- Supports text, image, and video inputs with text output and a context window of up to 1,000,000 tokens.
- Scores 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.
- Uses MiniMax Sparse Attention throughout the model, quantization-aware training, and DSpark speculative decoding.
- Released under the minimax-community license; library_name is transformers.
Why it matters
The model pairs a 1M-token context with sparse attention and speculative decoding aimed at inference efficiency, targeting agentic, coding, and cybersecurity tasks. Its benchmark table positions it against DeepSeek V4.1 Flash, Kimi K3, GLM-5.3, GPT-5.6 Sol, and Opus 5.