Xiaomi Launches MiMo-V2.6: 3D and API Access

Xiaomi has made MiMo-V2.6 Pro and MiMo-V2.6-Flash available as open source, with distribution on HuggingFace, OpenRouter, and via API at the same price as V2.5 Pro. The models, which can run locally on smartphones, claim a high rank on a benchmark and introduce a 3D feature capable of generating interactive environments.

Open Source Access, Platforms, and Desktop Applications

Xiaomi is releasing the weights of its MiMo-V2.6 models as open source, accompanied by over 7,000 reinforcement learning environments and its complete RL infrastructure. The models are available on HuggingFace, OpenRouter, and via API at a price identical to that of the V2.5 Pro version. At the same time, the company is expanding the deployment of its MiMo desktop applications for Windows and macOS. Both models can run locally on Xiaomi smartphones.

Omnimodal Capabilities and 3D Universe Generation

MiMo-V2.6 Pro and MiMo-V2.6-Flash are omnimodal and handle text, image, video, and code. They incorporate a 3D spatial reasoning feature, Vibe World, capable of producing interactive environments, including video games, from instructions in the form of text, image, or video. Announced use cases include generating models for Blender, composing demo-format tracks, and controlling a Franka Panda robotic arm from a multi-camera stream.

Claimed Performance and Positioning of the Two Variants

MiMo-V2.6 Pro is presented as the most powerful, while MiMo-V2.6-Flash aims for faster responses and reduced usage costs. According to Xiaomi, MiMo-V2.6 Pro ranks at the top of the open weights models on the Artificial Analysis Intelligence Index benchmark, ahead of Alibaba's Qwen 3.8 Max. Xiaomi also claims that the model is slightly behind Claude Opus 5, GPT-5.6 Sol, and Meta Muse Spark, and not far from Claude Fable 5.1, while being well ahead of DeepSeek Flash Max.