Xiaomi Releases MiMo-V2.6-Flash-RL, a 309B-Parameter MoE Model with 1M-Token Context

Xiaomi's MiMo team has released MiMo-V2.6-Flash-RL, the efficiency-balanced checkpoint of its MiMo-V2.6 model series. The model is a Sparse Mixture-of-Experts (MoE) architecture with 309 billion total parameters and 15 billion active parameters per forward pass, and it supports a 1 million token context window.

The release is available now on Hugging Face and ModelScope, alongside a larger sibling, MiMo-V2.6-Pro-RL. Deployment is also supported through Xiaomi's AI Studio, MiMo Code, MiMo Desktop, the Xiaomi MiMo Open Platform API, and OpenRouter. Pricing has not yet been disclosed.

Architecture and Modalities

MiMo-V2.6-Flash-RL is natively omnimodal, processing text, image, video, and audio within a single model. It pairs a 48-layer LLM backbone (39 sliding-window attention layers, 9 global attention layers, 4096 hidden size, 256 routed experts with 8 activated) with a 681-million-parameter MiMo ViT vision encoder (28 layers: 24 sliding-window, 4 full attention) and a two-part audio stack — a 308M-parameter AudioTokenizer using 20 RVQ codebooks and a 127M-parameter audio patch encoder. A 5-layer speculative decoder (DFlash-style multi-token prediction) predicts seven tokens per forward pass for faster inference.

According to Xiaomi, the model is trained with a single mixed reinforcement learning run — dubbed "You Only RL Once" — spanning coding, general agent tasks, visual reasoning, and cybersecurity simultaneously, rather than separate per-domain training runs. The company claims this cross-domain mixing lets capabilities reinforce each other and generalize to agent harnesses not seen during training.

The training approach also introduces what Xiaomi calls Groupwise Agentic Grading, which replaces binary pass/fail rewards with an agentic grader that ranks rollouts within each group using two mechanisms: Groupwise Reward Synthesis, which builds task-specific rubrics from contrasting rollouts, and Groupwise Advantage Redistribution, which shifts advantage toward higher-quality trajectories online. Xiaomi says this closes a self-improvement loop that also favors shorter, more token-efficient solutions. RL is run at large scale — 1,568 prompts by 16 rollouts per step — using fully asynchronous Group Relative Policy Optimization (GRPO).

Benchmark Results

According to figures published by Xiaomi, MiMo-V2.6-Flash-RL scores 67.9 on DeepSWE v1.1, 61.2 on MiMo Code Bench, 52.3 on AutomationBench v1.0.6, 87.6 on Terminal Bench 2.1, and 95.1 on CyberGym. The company's benchmark tables place these results alongside comparison figures for Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5 — for example, Terminal Bench 2.1 scores of 89.1 (Claude Opus 5), 88.8 (GPT-5.6 Sol), and 84.3 (Claude Fable 5). These comparison numbers come from Xiaomi's own technical report and have not been independently verified.

The smaller Flash checkpoint trails the larger MiMo-V2.6-Pro-RL on most benchmarks — for instance 82.0 vs. 80.8 on OSWorld-Verified and 94.0 vs. 95.1 on CyberGym — reflecting the expected tradeoff for reduced active parameter count.

Deployment

Xiaomi recommends serving the model through SGLang or vLLM, both of which have MiMo-specific recipes and pre-built Docker images. Recommended sampling settings are temperature 1.0 and top_p 0.95. Note that the Hugging Face repository lists the safetensors file size as 159B parameters, which does not match the 309B total parameter figure stated in the model card — this discrepancy has not been clarified by Xiaomi.

What This Means

MiMo-V2.6-Flash-RL extends Xiaomi's push into large-scale, omnimodal, RL-trained agentic models, competing directly with frontier labs on coding, tool-use, and cybersecurity benchmarks. The unified RL training approach across domains is notable methodologically, but the benchmark comparisons against unreleased or unverified competitor models (Claude Opus 5, GPT-5.6 Sol) should be treated as Xiaomi's own claims until independently reproduced. The unexplained parameter-count discrepancy between the model card (309B) and the hosted file size (159B) is worth watching as the model sees wider deployment and community verification.