Xiaomi MiMo released MiMo-V2.6-Pro-MOPD on Hugging Face
Model: MiMo-V2.6-Pro · availability, license and releases
Xiaomi MiMo released MiMo-V2.6-Pro-MOPD on Hugging Face, an upgraded checkpoint of MiMo-V2.6-Pro-RL that mitigates tool-call repetition and extends to domains like game development, scientific research, and embodied intelligence. It is a sparse MoE model with 1.02T total parameters, 42B activated, 1M context length, and MIT license.
China context
- Original name
- 小米 MiMo
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can download the MIT-licensed weights from Hugging Face and integrate the model into agentic applications, with the expectation of reduced tool-call repetition compared to the previous RL checkpoint.
- For investors
- Investors should note Xiaomi's continued investment in open-weights multimodal models with agentic focus, which may pressure other open-weights providers to improve reliability in tool use.
Xiaomi MiMo released MiMo-V2.6-Pro-MOPD on Hugging Face. This checkpoint is an MOPD upgrade of MiMo-V2.6-Pro-RL, fusing several domain-specialized teachers into one model to extend to domains where reliable training-time verification is hard, such as long-horizon game development, scientific research, and embodied intelligence. It also mitigates tool-call repetition, a failure mode where the model issues the same or highly similar tool calls repeatedly. The model is a sparse Mixture of Experts (MoE) with 1.02T total parameters and 42B activated parameters, supporting a 1M token context length and text, image, video, and audio modalities. It is available on Hugging Face and ModelScope under the MIT license.
MiMo-V2.6-Pro-MOPD uses MOPD2, which distills several domain-specialized teachers into the student on-policy. Teachers include mixRL teachers trained on verifiable tasks and SFT teachers trained on synthetic demonstrations for open-domain tasks. Three streams contribute to a single update: Standard MOPD, Teacher-Prefix OPD, and SFT-Prefix OPD. The model architecture is a sparse MoE with 70 layers (60 SWA, 10 GA), hidden size 6144, 384 routed experts with 8 activated, and a 5-layer speculative decoder. The vision encoder is a 681M-param MiMo ViT with 28 layers, and audio encoders include a 308M AudioTokenizer and a 127M audio patch encoder.
Developers using Xiaomi MiMo for agentic applications can now access a checkpoint that reduces tool-call repetition, which can lower token costs and improve task completion rates in long-horizon tasks. The open-weights release under MIT license allows competitors and researchers to inspect and build upon the model, increasing pressure on other open-weights providers to address similar agentic reliability issues.
The release of MiMo-V2.6-Pro-MOPD under MIT license enables enterprises and developers to deploy a state-of-the-art multimodal model with improved agentic reliability without licensing fees. The mitigation of tool-call repetition can reduce operational costs in agentic workflows by avoiding wasted API calls and context usage.
Observable next signals include adoption metrics on Hugging Face (downloads, likes), community feedback on tool-call repetition in agentic benchmarks, and whether Xiaomi releases a technical report or blog post detailing the MOPD2 method and its effectiveness. Also watch for integration into Xiaomi's API platform and other deployment channels.