Skip to content
AI.info

The Pulse

Xiaomi Open-Sources MiMo-V2.6 and Its Reinforcement-Learning Stack

Xiaomi announced MiMo-V2.6 on September 22, releasing Pro and Flash model weights alongside reinforcement-learning code, training environments and a technical report. The company’s six-day training run completed about 750,000 trajectories, while its published benchmark results show strengths and gaps against leading closed models.

Xiaomi Open-Sources MiMo-V2.6 and Its Reinforcement-Learning Stack

Xiaomi’s MiMo-V2.6-Pro performs near leading closed models on many agent tests, the company says, but its own benchmark table shows results that vary by task. The September 22 release pairs Pro and Flash model weights with a larger package: Xiaomi’s reinforcement-learning code, training environments and technical report.

The announcement presents the release as a bid to make large-scale agent training reproducible, not just to publish another set of model weights. Xiaomi’s two models completed 30 training steps each in less than six days, across roughly 750,000 trajectories. Its claims and results are company-reported; the published materials do not establish independent replication.

Thirty training steps, 750,000 trajectories

Xiaomi spent about $2.62 million training MiMo-V2.6-Pro and about $850,000 on Flash. Average training-task pass rates rose 12% for Pro and 25% for Flash, the firm reports. On its out-of-sample DeepSWE v1.1 software-engineering evaluation, Pro moved from 58.4 to 72.6 and Flash from 48.8 to 65.7, Xiaomi reports.

The training setup scales three elements together: batch size and throughput, task diversity, and the computing power used by graders to assess model outputs. Each update used 1,568 samples, with training supporting a one-million-token context and processing 3.5 billion to 3.7 billion tokens per step. Tasks covered coding, general agent work, visual tasks and cybersecurity.

The open release includes environments and training code

Xiaomi lists more than 7,000 reinforcement-learning task environments across software engineering, vulnerability reproduction, knowledge-intensive work, and web design and development. It also released an end-to-end training framework built on verl, uni-agent and mini-swe-agent, covering environment interaction, trajectory collection, reward evaluation and policy optimization.

The release includes MiMo-V2.6-Pro and Flash weights and technical reports, plus MiMo-V2.6-Distill-Qwen-9B and supporting RL research resources. The training process freezes the mixture-of-experts router to limit expert-load drift and uses reward design, adversarial evaluation, anomaly detection and validator cross-checks to address reward hacking. The Pro model card lists the model under the MIT license; that label should not be assumed to apply automatically to every element of the wider training stack.

“We have open-sourced the aforementioned technical achievements and supporting resources, including the complete technical report, training environment and RL code, to help more researchers reproduce and verify relevant results, and jointly explore more possibilities of large-scale RL and model self-improvement.”

Xiaomi MiMo

Benchmark results vary by task

Xiaomi describes Pro as matching Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks. Its published Pro model card gives a more detailed comparison: Pro scores 71.9 on DeepSWE v1.1, below the listed 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol. On Toolathlon-Verified, Pro scores 76.9 against Opus 5’s 80.6, while on AutomationBench v1.0.6 it scores 53.1, above Opus 5’s 50.3.

Those figures come from Xiaomi’s model card, not an independent evaluation, and results on one test do not establish broad superiority. The release gives researchers materials to inspect and try to reproduce the training, but the company’s reported costs, pass-rate gains and benchmark comparisons remain claims made by Xiaomi.

Weights and API access arrive together

The new models keep the API pricing of its V2.5 series, while adding MiMo-V2.6-Pro UltraSpeed, which the company says can deliver up to 20 times faster inference. The models are also available through Xiaomi’s API platform and desktop client. The release’s central contribution is the public training package: alongside the models, it exposes the environments and code Xiaomi says it used to train them.

Sources

  1. Original source Xiaomi MiMo mimo.mi.com

Other pages read

  1. Hugging Face huggingface.co

Related stories

Explore

More articles