MiniMax has released M3.1 Flash, the newest model in its lower-cost Flash line, and used the launch on Monday, October 19, 2026 to clear up a small mystery in the developer community. Space Bunny, a model that had been circulating without a named owner, was an early version of M3.1 Flash, the Shanghai-based company said, adding that it processed more than 200 trillion tokens while operating anonymously. Reuters reported the launch the same day, noting that MiniMax claims major improvements.
The company is pitching the model at agent builders. MiniMax describes M3.1 Flash as a generational leap in agentic tasks, coding, computer use and cybersecurity, and says it completes real enterprise workflows across software-as-a-service tools in sales, marketing, operations, customer service, finance and HR. Its own benchmark table lists 73.7% on DeepSWE v1.1 for software engineering, 76.2 on BankerToolBench, 61.7 on OSWorld V2 for computer operation, 29.6 on Agents' Last Exam and 82% on CyberGym for vulnerability discovery. MiniMax says the OSWorld V2 and Agents' Last Exam results rank first among Chinese models.
Speed and price are the other half of the pitch. MiniMax says more aggressive sparsity makes online inference nearly three times faster than before, with output often above 200 tokens per second and a median of more than 170. It says the overall cost of agentic tasks is 70% lower than with the previous Flash model. List pricing is $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens.
The model is available now through the MiniMax API, the M Plan subscription and the MiniMax Code desktop agent. Weights have not been published: the company says they will be released later, without giving a date, and that a larger Pro model will follow. Until the weights appear, developers can use M3.1 Flash only as a hosted service.
The Space Bunny detail gives the launch some outside grounding. TokenPost reported that the preview appeared on OpenRouter as Space Bunny Alpha on September 23 and held the top spot in OpenCode's daily token-usage ranking by September 24. Usage of that kind reflects developer interest and price rather than quality, but it suggests the model had already been tested on real coding workloads before it carried MiniMax's name.
The release lands in a crowded week for Chinese budget models. Two days earlier DeepSeek published DeepSeek-V4.1-Flash-1017, an open-weight update that it says scores 77.0 on DeepSWE v1.1, above the 73.7 MiniMax reports for M3.1 Flash. Both figures are vendor numbers, harness settings may differ, and neither has been reproduced independently, so the comparison is indicative at best. What is clear is that Flash-tier models are now competing on agentic coding scores and cost per task rather than on chat quality.
For teams running agents at volume, the main draws are the price and the stated speed. The main caveats are that every capability figure in the announcement comes from MiniMax itself, that the cost reduction is measured against MiniMax's own predecessor, and that the open release is still a promise. Independent evaluations, and the timing of the weights, will decide how far the launch changes developers' choices.