MiMo-V2.6 Pro

// Overview

MiMo-V2.6 Pro is the flagship checkpoint of Xiaomi's MiMo-V2.6 series, published to Hugging Face as MiMo-V2.6-Pro-RL on 21 September 2026 under an MIT licence. The model card describes a sparse Mixture-of-Experts model with 1.02 trillion total parameters and 42 billion activated per token, a 1 million-token context, and native handling of text, image, video and audio in a single model.

The series' stated goal is to "scale reinforcement learning toward self-improvement" — scaling RL compute, environment diversity and grader compute together. Xiaomi describes a single mixed RL run ("You Only RL Once") that trains coding, general-agent, visual and cybersecurity tasks in the same batch across multiple harnesses, using fully asynchronous GRPO on batches of 1,568 prompts × 16 rollouts per step.

The reward signal is itself scaled through what the card calls Groupwise Agentic Grading. Groupwise Reward Synthesis builds task-specific rubrics offline from contrasting rollouts and fuses rubric quality with test outcomes; Groupwise Advantage Redistribution ranks passing trajectories online and shifts advantage toward higher-quality solutions. A later stage, Multi-Prefix Multi-Teacher On-Policy Distillation, reuses teacher and SFT histories so decision points can be trained without regenerating the preceding turns.

Architecturally the backbone interleaves local sliding-window attention with global attention and uses sparse MoE feed-forward blocks with no shared experts. Omnimodal input arrives through a 681M-parameter MiMo ViT and a two-part audio stack, and a five-layer speculative decoder predicts seven subsequent tokens per forward pass. Weights, tokenizer, model card and a technical report are all published; deployment recipes are given for SGLang and vLLM.

// Benchmarks

This model's scores

1. CyberGym94%
2. Terminal Bench 2.189.9%
3. OSWorld-Verified82%
4. MiMo Cyber Bench80.2%
5. Toolathlon-Verified76.9%
6. MiMo VisualCoding72.3%
7. DeepSWE v1.171.9%
8. SEC Bench Pro66.3%
9. MiMo Code Bench63.2%
10. JobBench62%
11. AutomationBench v1.0.653.1%
12. ExploitBench47.9%
13. Terminal Bench 4.034.9%
14. Agents' Last Exam31.6%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• MIT licence on a 1.02T-parameter frontier-class MoE — weights, tokenizer and technical report are all downloadable
• Native omnimodal input (text, image, video and audio) in one model, with a 1M-token context for long repositories and multi-session agent runs
• Leads its own published comparison on AutomationBench v1.0.6 (53.1), Terminal Bench 2.1 (89.9) and CyberGym (94.0)
• A single mixed reinforcement-learning run across coding, agent, visual and cybersecurity tasks rather than separate per-domain runs
• Documented SGLang and vLLM serving paths, including EAGLE speculative decoding and MiMo-specific reasoning and tool-call parsers

// Best for

• Reach for it when an agent stack needs frontier-class coding and tool use on weights you can host yourself under a permissive licence.
• Reach for it for long-horizon work — large repositories, long tool traces, multi-session runs — that needs the full 1M-token context.
• Reach for it when a workflow mixes modalities: reading screenshots or video alongside code and text in the same conversation.
• Reach for it for authorised security work, where the card reports its strongest margins over the previous generation.

// How to access

// MiMo (open weights) — every version

The full lineage of the MiMo (open weights) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.