On HuggingFace, on September 21, 2026, the model XiaomiMiMo/MiMo-V2.6-Flash-RL was published with the text-generation task. The author describes it as an efficiency-balanced checkpoint of the MiMo-V2.6 series, built around scaling reinforcement learning toward self-improvement: a single mixed RL run across coding, agents, visual tasks and cybersecurity, asynchronous GRPO on large batches, and group agentic evaluation instead of binary checking. Details are on the model page.
What it means
The model is large: 159.4 billion parameters, MoE architecture. The full weights in the repository take up 165.5 GB, the largest file is 12.5 GB. License — mit. It runs via transformers, the weights are posted in safetensors and pt.
For local running, the site has 24 quantized builds: FP8, MLX, GGUF, NVFP4, GPTQ and EXL. The smallest working build is AesSedai/MiMo-V2.6-Flash-GGUF at 90.1 GB, not counting variants worse than 4 bit. That is still a substantial volume: loading such a build requires a corresponding amount of system RAM or VRAM, and the developer does not specify minimum hardware requirements.
You can follow the arrival of new AI tools for 3D and video and the models that run on your own hardware in the CPU3D news feed.How the method works. The diagram was drawn based on this news note.
