MiniMax open-sources M3.1 Flash multimodal MoE weights
MiniMax released the open weights for M3.1 Flash, a multimodal mixture-of-experts model with 428B total parameters and 23B activated per token, according to IT之家 AI and MiniMax’s Hugging Face model card. It accepts text, image, and video input with a context window of up to 1M tokens, while MiniMax says sparse attention, quantization-aware training, and power-aware scheduling nearly tripled inference speed and reduced agent-task costs by 70% versus M3. The weights are available through Hugging Face and ModelScope, with Tenstorrent Cloud, vLLM, and SGLang among the platforms and frameworks cited as supporting deployment.
The release gives developers direct access to a large multimodal MoE model built for coding, agentic workflows, computer use, and cybersecurity, while its day-one hardware and serving support broadens the options for self-deployment.
MiniMaxAI/MiniMax-M3.1-Flash Image-Text-to-Text • 220B • Updated 17 days ago • 151k • • 610
pipeline_tag: image-text-to-text license: other license_name: minimax-community license_link: LICENSE library_name: transformers tags:
- multimodal
- moe
- agent
- coding
- video
#MiniMax#MiniMax M3.1 Flash#Tenstorrent Cloud#vLLM#SGLang#Hugging Face#ModelScope