Xiaomi Releases MiMo-V2.6-Pro-RL, a 1.02T-Parameter Omnimodal Model with 1M-Token Context

Xiaomi's MiMo team has released MiMo-V2.6-Pro-RL, the flagship checkpoint of its new MiMo-V2.6 series, on Hugging Face. The model is a sparse Mixture-of-Experts (MoE) architecture with 1.02 trillion total parameters and 42 billion activated per forward pass, supporting a 1 million token context window and native processing of text, image, video, and audio in a single model.

Architecture

The LLM backbone has 70 total layers (60 sliding-window attention, 10 global attention), a hidden size of 6144, and 384 routed experts with 8 activated per token. A vision encoder (MiMo ViT, 681M parameters, 28 layers) and audio stack (a 308M-parameter AudioTokenizer plus a 127M-parameter audio patch encoder) handle multimodal inputs. A 5-layer multi-token prediction module serves as a speculative decoder, predicting seven tokens per forward pass for parallel verification.

According to Xiaomi, the model was trained using a single mixed reinforcement learning run — dubbed "You Only RL Once" — combining coding, general agent, visual, and cybersecurity tasks in the same training batches rather than running separate domain-specific RL passes. The company says this was done at large scale: fully asynchronous Group Relative Policy Optimization (GRPO) with 1,568 prompts × 16 rollouts per step, processing billions of tokens per update.

Xiaomi also describes a custom reward mechanism it calls Groupwise Agentic Grading, which uses an agentic grader to rank passing rollouts within each group rather than relying on binary pass/fail signals, plus a distillation technique (MOPD2) that reuses trajectory histories from teacher and SFT demonstrations to extend training to harder-to-verify tasks.

Benchmark Results

Xiaomi's published evaluation table compares MiMo-V2.6-Pro against MiMo-V2.6-Flash, MiMo-V2.5-Pro, and unspecified competitor models labeled "Claude Opus 5," "GPT-5.6 Sol," and "Claude Fable 5" — none of which are confirmed public releases from Anthropic or OpenAI as of this writing, so these comparisons should be treated as claims from Xiaomi's technical report rather than independently verified figures.

Reported scores for MiMo-V2.6-Pro include: DeepSWE v1.1 71.9, ProgramBench 26.5, MiMo Code Bench 63.2, AutomationBench v1.0.6 53.1, Toolathlon-Verified 76.9, Agents' Last Exam 31.6, Terminal Bench 4.0 34.9, Terminal Bench 2.1 89.9, OSWorld-Verified 82.0, JobBench 62.0, CyberGym 94.0, MiMo Cyber Bench 80.2, ExploitGym 17.8, ExploitBench 47.9, SEC Bench Pro 66.3, and MiMo VisualCoding 72.3. These represent large jumps over MiMo-V2.5-Pro on several tasks — for example, MiMo Cyber Bench rose from 0.0 to 80.2.

Availability

The model is available on Hugging Face and ModelScope, with deployment recipes published for SGLang and vLLM. Xiaomi lists additional access points including its AI Studio, MiMo Code, MiMo Desktop app, the Xiaomi MiMo Open Platform API, and OpenRouter. Pricing for API access has not yet been disclosed. A smaller companion model, MiMo-V2.6-Flash-RL, was released alongside it.

What This Means

MiMo-V2.6-Pro-RL signals Xiaomi's continued push into frontier-scale foundation models, competing directly on agentic coding and cybersecurity benchmarks — domains typically dominated by US labs. The single mixed-RL-run approach and 1M-token context place it in the same tier as top-end omnimodal systems, though the benchmark comparisons against models like "GPT-5.6" and "Claude Opus 5" cannot be independently confirmed and should be read as Xiaomi's own reported figures until third-party evaluations appear.