MiniMax releases open weights for M3.1 Flash AI model
MiniMax has released the weights for M3.1 Flash, a multimodal 428-billion-parameter model aimed at coding, agents and computer use.
MiniMax has released the weights for its M3.1 Flash model, giving developers access to a 428-billion-parameter mixture-of-experts system with 23 billion parameters activated per token. The model accepts text, image and video inputs and supports a context window of up to one million tokens, according to its official model page.1
What the open release includes
The M3.1 Flash release is aimed at coding, AI agents, computer-use tasks and cybersecurity. MiniMax’s model documentation lists sparse attention, quantisation-aware training and speculative decoding among the techniques used to reduce the amount of computation needed during inference.1
The model page also provides deployment information for developers, while MiniMax’s official announcement describes the model’s intended use in coding, computer use and agentic workflows. That makes this more significant than another hosted model launch: researchers and engineers can inspect the files, adapt the deployment and run the model on their own infrastructure where the hardware and licence allow it.12
MiniMax’s own announcement presents M3.1 Flash as a faster and cheaper model for agentic workflows, coding and computer use. Its published comparison table reports a 73.7 score on DeepSWE v1.1, 61.7 on OSWorld 2.0 and 82.2 on CyberGym, although those figures are company-reported and should not be treated as independent confirmation of real-world performance.2
Why it matters
Open weights do not make a 428-billion-parameter model easy to run. The active-parameter figure reduces the work done per token, but deployment still demands serious memory, bandwidth and systems engineering. The practical audience is therefore likely to be research teams, cloud providers and well-funded developers rather than ordinary desktop users.
The release nevertheless broadens the options for organisations that want multimodal models under their own control. It also gives the wider research community a chance to test MiniMax’s claims against reproducible workloads rather than relying solely on an API demo.
Our opinion
The interesting part is not the headline parameter count; it is the combination of open access and an emphasis on computer-use tasks. If independent testing confirms the reported efficiency, M3.1 Flash could become a useful building block for agents that need to operate software rather than merely describe it. Until those tests arrive, the sensible view is promising release, unproven advantage.