MiniMax has released M3.1 Flash for coding and multi-step AI-agent work. The company identified the previously anonymous Space Bunny model as an early version. Access is available through its API, M Plan and MiniMax Code, with downloadable model weights promised later. The announcement includes company-supplied performance results, while the release itself is also reported by The Decoder.
MiniMax formally released M3.1 Flash on October 19 for coding and multi-step AI-agent tasks. The release is also reported by The Decoder. The company lists its API, M Plan subscription and MiniMax Code desktop application as access channels. Downloadable weights are due later, so the launch provides hosted access rather than a model that customers can already download and run locally.[1], [2]
The company identified the previously anonymous Space Bunny model as an early version of M3.1 Flash. It says that preview processed more than 200 trillion tokens. MiniMax’s examples of intended work include sales, marketing, operations, customer support, finance and human resources. The model is aimed at chains of tool use, desktop or browser interaction and software work, with cybersecurity tasks also included in the company’s evaluations.[1]
MiniMax reports 73.7 percent on DeepSWE v1.1 and 82 percent on CyberGym, alongside results from other task collections. These are company-supplied measurements under differing evaluation conditions. A larger Pro version is described as a later product, and the announcement leaves the distribution of model weights to a separate future stage.[1]
