GitHub radar
Xiaomi MiMo V2.6: MIT-licensed 309B agent model
Xiaomi released MiMo-V2.6-Flash-RL, a 309 billion parameter model for agentic tasks — it understands text, images, video, and audio, under the MIT license.
MiMo-V2.6-Flash-RL is, in the words of its authors, 'the efficiency-balanced checkpoint of the MiMo-V2.6 series.' The model from XiaomiMiMo — Xiaomi's AI team — processes text, images, video, and audio in a single request. It is designed for agentic tasks: it can write code on its own, operate a web browser, and interpret documents from their images. The context window is 1 million tokens. In total the model has 309 billion parameters, but the architecture activates only about 15 billion of them at a time. The license is MIT, which means any cloud provider can deploy it and offer API access without restrictions.
Why a vibe-coder should care
Regarding open models at this scale — Xiaomi releasing a 309B omnimodal model under MIT is not something that happens often. At the moment this model requires server hardware to run, but with MIT license it will appear in cloud APIs. For building with agents that need to work with different data types, this is a substantial option in the open ecosystem.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Show me how to try Xiaomi's MiMo model: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL — find where I can make a request through an API or demo right now, and show me a simple example with an image.
Server hardware required — nearly 310 billion parameters cannot run on a personal computer. This is a data center model.
Open on Hugging Face▌ More finds