The release is presented as a native multimodal mixture-of-experts model for coding, agentic workflows, computer use, and cybersecurity tasks.
According to the model card, M3.1-Flash accepts text, image, and video inputs and supports up to one million tokens of context. MiniMax says the model builds on MiniMax-M3 and uses sparse attention, quantization-aware training, and DSpark speculative decoding.
The card reports developer results of 76.2 on BankerToolBench and 73.7% on DeepSWE v1.1. Those figures are MiniMax's reported evaluations, not independent results. The source does not establish inference speed, hardware requirements, deployment cost, performance under a particular quantization setting, or whether the licence terms fit a given commercial use case.
The public weights make the model available for inspection and self-hosted evaluation subject to its stated licence.