MiMo-V2.6

MiMo-V2.6 is a series of natively omnimodal language models developed by the Xiaomi MiMo team and released on 21 September 2026 under the MIT license. The models take text, image, video and audio as input, generate text (including code) and support a context window of up to 1 million tokens.

The series comprises three variants: the flagship MiMo-V2.6-Pro-RL (sparse Mixture-of-Experts, 1.02T total / 42B activated parameters, 384 routed experts with 8 active), the efficiency-balanced MiMo-V2.6-Flash-RL (309B / 15B activated) and the distilled MiMo-V2.6-Distill-Qwen-9B (9B, SFT on top of Qwen3.5-9B). The architecture combines a hybrid sliding-window-attention (SWA) backbone, a MiMo ViT vision encoder (681M), audio encoders and a 5-layer speculative decoder (Multi-Token Prediction).

The models are trained via scaled reinforcement learning (Group Relative Policy Optimization) in a single mixed RL run spanning coding, general agents, visual tasks and cybersecurity, complemented by on-policy distillation (MOPD2) and a self-improvement loop (Groupwise Reward Synthesis, Groupwise Advantage Redistribution). The models support tool calling and agentic operation.