MiMo-V2.6-Pro is the flagship of Xiaomi’s September 22, 2026 MiMo-V2.6 release, and its pitch is unusual: it is a 1.02T-parameter, 42B-active, omni-modal model whose main improvement came from six days of large-scale reinforcement learning, with the training run itself livestreamed and then open-sourced. Xiaomi puts the compute bill at roughly $2.62M for Pro and $850K for Flash.
The result, by Xiaomi’s account, is an open model that scores 46 on the Artificial Analysis Intelligence Index, ahead of Kimi K3 and Qwen3.8-Max, and that matches Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks — at $0.435 per million input tokens. Weights are MIT-licensed, and Xiaomi also released 7,000+ RL task environments, an RL training framework and a 9B distilled model for the community.
Benchmarks
Vendor-reported by Xiaomi (September 22, 2026), as published on the model card.
| Benchmark | MiMo-V2.6-Pro | MiMo-V2.6-Flash | MiMo-V2.5-Pro | Claude Opus 5 | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|---|---|---|---|
| DeepSWE v1.1 | 71.9 | 67.9 | 19.0 | 74.0 | 73.0 | 70.0 |
| ProgramBench | 26.5 | 26.0 | 12.5 | 37.0 | 25.0 | 33.0 |
| AutomationBench | 53.1 | 52.3 | 16.0 | 50.3 | 45.8 | 46.2 |
| Toolathlon Verified | 76.9 | 73.6 | 49.1 | 80.6 | 74.9 | 77.9 |
| Agents’ Last Exam | 31.6 | 27.6 | 13.2 | 31.6 | 30.8 | 25.7 |
| Terminal-Bench 2.1 | 89.9 | 87.6 | 65.2 | 89.1 | 88.8 | 84.3 |
| Terminal-Bench 4.0 | 34.9 | 28.8 | 1.5 | 49.0 | 39.9 | 42.4 |
| OSWorld-Verified | 82.0 | 80.8 | — | 83.4 | 83.0 | 86.0 |
| GDPval-AA 2.1 (Elo) | 1673 | — | 1107 | 1708 | 1588 | 1595 |
| CyberGym | 94.0 | 95.1 | 40.0 | — | — | — |
| ExploitBench | 47.9 | 25.3 | 16.6 | 70.0 | 78.5 | 78.0 |
| MiMo VisualCoding | 72.3 | 71.5 | — | 70.0 | 73.4 | 69.1 |
Two notes. The V2.5-Pro column (19.0 on DeepSWE v1.1) and Xiaomi’s blog (58.4 rising to 72.6 over the six-day run) report different baselines, so treat the size of the generational jump with caution. And the cyber rows are the weakest comparison — closed-model scores are missing on several and Xiaomi’s own MiMo Cyber Bench is vendor-defined.
How the RL worked
- One mixed run across coding, general agents, visual tasks and cybersecurity, with multiple harnesses in the same batch so skills transfer to frameworks unseen in training.
- Fully asynchronous GRPO with 1,568 prompts and 16 rollouts per step, at 1M context and 3.5–3.7B tokens per update.
- Groupwise agentic grading — a grader compares rollouts within each group, building task-specific rubrics and re-ranking passing solutions, which steers the model toward shorter paths and fewer tokens.
- Anti-reward-hacking measures and a frozen MoE router to stop expert load drifting.
Pricing
| Per 1M tokens | USD | CNY |
|---|---|---|
| Input (cache miss) | $0.435 | ¥3 |
| Input (cache hit) | $0.0036 | ¥0.025 |
| Output | $0.87 | ¥6 |
Rate limits are 100 requests and 10M tokens per minute. The Token Plan subscription (individual and team) covers mimo-v2.6-pro and mimo-v2.6-flash, and a Batch API is available for non-real-time jobs. mimo-v2.6-pro-ultraspeed serves up to 20x faster for latency-sensitive use.
Specs
| Spec | Value |
|---|---|
| API models | mimo-v2.6-pro, mimo-v2.6-flash, mimo-v2.6-pro-ultraspeed |
| Architecture | Sparse MoE, 1.02T total / 42B active; 70 layers (60 SWA, 10 global); 384 routed experts, 8 active; 5-layer MTP |
| Context window | 1M tokens; max output 128K |
| Modalities | Text, image, video, audio in; text out |
| Encoders | 681M-parameter MiMo ViT; audio tokenizer plus patch encoder |
| Weights | Hugging Face XiaomiMiMo/MiMo-V2.6-Pro-RL (BF16 / FP8) |
| License | MIT |
| Serving | SGLang, vLLM (cookbooks provided) |
How it compares
- vs Kimi K3 and Qwen3.8-Max: Xiaomi says it beats both on the Artificial Analysis index; both are open-weight rivals at a larger parameter count.
- vs GLM-5.3: both are MIT-style open models near the top of the open field; GLM-5.3 is text-only, MiMo is omni-modal.
- vs DeepSeek V4.1-Flash: DeepSeek’s DeepSWE v1.1 is slightly higher (74.2 vs 71.9) at a smaller active size; MiMo adds audio and video input.
- vs closed frontier: it sits just below Claude Opus 5 and GPT-5.6 Sol on most agent rows, and behind the newer Claude Opus 5.5 and GPT-6 generation.
Limitations
- Vendor-reported. Scores and the Artificial Analysis claim come from Xiaomi.
- Text-only output.
- Cyber exploitation lags the closed models.
- Heavy to self-host at 1.02T parameters.
For the wider picture, see the 2026 LLM guide.
Best for
- Long-horizon coding and computer-use agents at low cost
- Video and audio understanding combined with tool use
- Cybersecurity research in sandboxed environments
- Studying large-scale RL, with the released environments and training code
Pros & cons
Strengths
- Xiaomi puts it at 46 on the Artificial Analysis Intelligence Index, ahead of Kimi K3 and Qwen3.8-Max — the strongest open-weight model on that index at launch
- Native omni-modal input — text, image, video and audio in one 1.02T model — plus a 1M-token context
- Near-frontier agent scores at about $0.44 per 1M input tokens — DeepSWE v1.1 71.9, Terminal-Bench 2.1 89.9, AutomationBench 53.1 (above Claude Opus 5's 50.3)
- Everything is open — MIT weights for Pro and Flash, the technical report, 7,000+ RL task environments and the RL training code
Limitations
- Still behind the closed leaders — Xiaomi itself says a gap remains to Claude Fable 5.1 and GPT-6 Astra — and behind Opus 5 on Terminal-Bench 4.0 (34.9 vs 49.0)
- Weaker on the cyber exploitation rows — ExploitBench 47.9 vs 70.0–78.5 for the closed models
- Output is text only; speech synthesis is a separate MiMo-V2.5-TTS family
- A 1.02T-parameter MoE is a 16-GPU-class self-hosting job (Xiaomi's recipe uses tensor parallel 16 over two nodes)
Frequently asked questions
When was MiMo-V2.6-Pro released?
Xiaomi released the MiMo-V2.6 series on September 22, 2026, with open weights, a technical report and RL resources on Hugging Face, and the API on the Xiaomi MiMo Open Platform as `mimo-v2.6-pro`.
How much does MiMo-V2.6-Pro cost?
$0.435 per 1M cache-miss input tokens, $0.0036 per 1M cache-hit input tokens and $0.87 per 1M output tokens (¥3, ¥0.025 and ¥6). That is unchanged from the V2.5 series. Xiaomi says it is 1/20 to 1/60 of overseas models at a similar intelligence level.
How big is MiMo-V2.6-Pro?
It is a sparse mixture-of-experts with 1.02T total and 42B active parameters, 70 layers (60 sliding-window, 10 global attention), 384 routed experts with 8 active, a 681M-parameter vision encoder and an audio tokenizer, plus a five-layer multi-token-prediction head.
What is the difference between MiMo-V2.6-Pro and Flash?
Pro is the trillion-parameter flagship; Flash is the lower-cost, higher-frequency model. Xiaomi says Flash comprehensively outperforms last generation's V2.5-Pro. Both are open-weight, fully multimodal and use the same API pricing structure as V2.5.
What is changing for V2.5 users?
MiMo-V2.5-Pro and V2.5 requests auto-route to V2.6 on October 14, 2026 at 18:00 (UTC+8). The old model IDs become invalid on October 21 at 10:00 (UTC+8), so migrate before then.