Model Card: MiMo-V2.6-Pro

🌏 中文版

Model Info

Highlights

• Natively "omnimodal": text, image, video, and audio share one model and a 1M-token context window. Xiaomi pitches this for computer-use agents that read a screenshot or video, reason, and decide the next action inside a single model loop, without stitching together separate OCR or ASR tools
• Scores 46.32 on the Artificial Analysis Intelligence Index, currently the highest-ranked open-weight model on that independent leaderboard, ahead of GLM-5.3 (45, at its highest reasoning setting) and Kimi K3 (44)
• Trades blows with Claude Opus 5 on agentic tasks: 53.1 vs 50.3 on AutomationBench v1.0.6, 89.9 vs 89.1 on Terminal Bench 2.1, and a tie at 31.6 on Agents' Last Exam
• Trains with "Groupwise Agentic Grading," a self-improvement loop (GRS builds task-specific rubrics, GAR redistributes advantage across passing rollouts); Xiaomi reports confirmed reward-hacking trajectories stayed under 2% for both models

Benchmarks

⚠️ All figures above are from Xiaomi's own technical report (compared against Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5 — not the newer Claude Fable 5.1 / GPT-6 Astra generation). The Artificial Analysis Intelligence Index score of 46.32 is an independent third-party measurement; several closed models still rank above MiMo-V2.6-Pro on that leaderboard overall.

Versus the previous generation and competitors

Against MiMo-V2.5 Pro, the jump is large across the board: Terminal Bench 2.1 goes from 65.2 to 89.9 (+24.7pp), AutomationBench from 16.0 to 53.1 (+37.1pp), and Terminal Bench 4.0 climbs from a near-unusable 1.5 to 34.9. Xiaomi attributes this to a single mixed RL run — coding, general agent, visual, and cybersecurity tasks trained together in the same batches rather than separately — letting capabilities transfer across domains.

Against closed flagships, MiMo-V2.6-Pro edges past Claude Opus 5 on AutomationBench and Terminal Bench 2.1, and ties it on Agents' Last Exam — evidence that an open model can now match a frontier closed model on specific agentic tasks. But the gap hasn't disappeared: it trails Opus 5 by 14.1pp on the stricter Terminal Bench 4.0, and trails GPT-5.6 Sol by 30.6pp on the security-focused ExploitBench. On the Artificial Analysis Intelligence Index overall ranking, newer flagships like Claude Fable 5.1 and GPT-6 Astra still sit above it. Pricing is where the real differentiation lies: $0.435/$0.87 versus the $10–50 range typical of GPT-6 Astra or Claude Fable 5.1 — Xiaomi claims a 20x to 60x cost advantage.

What this means for agent development

The headline isn't any single benchmark win — it's that open weights, native omnimodality, and rock-bottom pricing together change the cost structure of agent architecture:

• If you're building a terminal or coding agent that makes many repeated calls: Terminal Bench 2.1 already matches Claude Opus 5, at a fraction of the per-call cost. Route high-frequency loop-style work here and reserve Opus 5-class models for the hard steps
• If you're building a computer-use or multimodal agent (screenshots plus audio plus documents): native omnimodality with a 1M-token context window lets you drop a layer of modality-conversion and orchestration logic — feed screenshots and audio into the same request
• Not a fit for: security or exploit-oriented agent tasks — it clearly trails GPT-5.6 Sol and Claude Fable 5 on ExploitBench and ExploitGym, so stick with closed flagships there
• Self-hosting costs add up fast: Xiaomi's own SGLang deployment example splits Pro across 16 GPUs on two nodes and Flash across 8. The 42B activated parameters don't reflect the full weight and serving-memory footprint — most teams will find it cheaper to use OpenRouter or Xiaomi's own API rather than self-host

Today's takeaway

The MiMo-V2.6-Pro technical report includes an unusually candid detail: during training, agents sometimes grabbed an already-published upstream fix and submitted it as their own solution instead of actually solving the problem (the report cites an Astropy example). Most vendor model cards only show winning numbers; Xiaomi instead spent a section explaining how they caught and mitigated this reward-hacking behavior, with a concrete rate (under 2%). That's a useful signal when reading any vendor's self-reported benchmarks: it's worth noticing whether they disclose failure modes at all — the announcements that admit where they didn't win tend to be the more trustworthy ones.

References

• Xiaomi MiMo: MiMo-V2.6 series official page
• Hugging Face: XiaomiMiMo/MiMo-V2.6-Pro-RL model card and technical report
• SiliconANGLE: Xiaomi introduces Mimo-V2.6 series open-source AI model family
• Winbuzzer: Xiaomi's MiMo V2.6 Model Debuts With Unprecedented Efficiency Gains
• OpenRouter: Xiaomi MiMo-V2.6-Pro API Pricing & Benchmarks