Xiaomi has released and open-sourced its MiMo-V2.6 family, turning a six-day public reinforcement-learning experiment into a broader bid for influence in advanced artificial intelligence. The release includes Pro and Flash, two natively multimodal systems designed to handle text, images, video and audio, as well as a smaller nine-billion-parameter checkpoint intended for reinforcement-learning research. Xiaomi also published model weights, a technical report, training code and more than seven thousand task environments. That combination matters because it gives outside developers more than a finished chatbot: it offers a set of components for testing how agentic systems can learn from repeated interaction with software, visual interfaces and other complex environments.
The company says each large model completed thirty reinforcement-learning steps and that the two runs produced roughly seven hundred fifty thousand trajectories. Xiaomi reported training costs of about eight hundred fifty thousand dollars for Flash and two million six hundred twenty thousand dollars for Pro, excluding the earlier work required to build the base systems. Its published results show gains on software-engineering, visual coding and cyber tasks. Independent testing adds some support, but also caution. Artificial Analysis scored Pro at forty-six on its intelligence index and placed it among the strongest open-weight systems it has tested. EWeek noted that this supports the finished model’s competitiveness without independently proving Xiaomi’s explanation of how the gains were achieved.
The release quickly reordered a Polymarket event asking which Chinese company will have the strongest artificial-intelligence model at the end of September. Xiaomi traded near eighty-eight percent Friday morning after gaining about eighty-seven percentage points over the prior day, while Alibaba and Moonshot AI were far behind. The selected Xiaomi market recorded about twenty-seven thousand five hundred dollars in latest-day activity and nearly eighty-eight thousand dollars overall. Those figures are trader-implied assessments, not benchmark verdicts, and the event’s resolution depends on its specified ranking source. Still, the speed of the move shows how decisively the open release changed expectations in a field that had recently favored established model developers.
The harder test begins after launch. Developers can now inspect the repositories, run the systems on their own workloads and compare vendor claims with repeatable evaluations. Xiaomi says Pro and Flash use MIT licenses and that the research package includes a distilled model, an end-to-end training framework and lightweight harnesses for building new tasks. Prospective users will also have to examine infrastructure demands, dependency licenses, security behavior and data controls. Benchmark leaders can change quickly, and several software-engineering tests have known limits involving contamination and scoring. MiMo-V2.6 therefore enters the competition with an unusually complete public package, but its lasting significance will depend on whether outside teams can reproduce useful gains rather than merely admire a launch-week leaderboard.
