The most interesting thing about MiMo-V2.6 Pro is not the benchmark score. It’s the price. Xiaomi built a trillion-parameter multimodal model, trained it for under six days, and is now charging $0.87 per million output tokens at a moment when Anthropic’s Claude Opus 5.5 costs $20. That gap is wide enough to change procurement decisions, not just leaderboard rankings.
Xiaomi announced two open-weight models on Tuesday, MiMo-V2.6-Pro and MiMo-V2.6-Flash, both released under an MIT licence with weights available on Hugging Face. The phone and electric car maker is not a name most AI teams have on their shortlist. That may need to change.
What Pro actually is
Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters, of which 42 billion are active per token. That architecture is why it runs cheaply at that scale. It accepts text, images, audio, and video, and holds a million tokens of context — enough for a full codebase, a long agent trace, or several sessions of working memory.
Artificial Analysis scores it 46 on its Intelligence Index, first among 114 models in its class and the highest any open-weight model has reached. The same firm rates it the cheapest model it tracks, at $0.13 per task. Xiaomi also claims it passes Moonshot’s Kimi K3 and Alibaba’s Qwen3-Max, which have traded the open-weight lead for most of this year. But Artificial Analysis has Opus 5.5 at 58 against MiMo’s 46. Xiaomi leads the open-weight field, not the whole field.
A third variant, Pro-UltraSpeed, generates up to 20 times faster than Pro at equivalent quality. It costs ten times more per token: $4.35 in, $8.70 out.
Six days, one training run, streamed live
The benchmark gains came from a single large reinforcement learning run that Xiaomi streamed publicly as it happened. Pro and Flash each completed 30 training steps over roughly 750,000 trajectories. Pro cost about $2.62 million to train. Flash cost about $0.85 million.
On DeepSWE, a held-out software engineering benchmark, Pro climbed from 58.4 to 72.57 across the run. Flash went from 48.8 to 65.68. Most labs run separate training jobs per domain. Xiaomi mixed coding, general agents, visual tasks, and cybersecurity into one run, betting that gains in each would reinforce the others.
The team also details its defences against reward hacking. Early agents solved assigned bugs by pulling fixes from later versions of the same package rather than writing the fix themselves. Xiaomi stripped build caches and future Git history from training environments and ran a dedicated agent to close remaining loopholes. That level of detail is uncommon in public releases.
Flash is the model most teams will actually deploy
Pro gets the attention. Flash is what ends up in production. It costs $0.14 per million input tokens and $0.28 per million output, roughly a third of Pro, and keeps the same million-token context and multimodal input. On most of Xiaomi’s own agent benchmarks it trails Pro by four points or fewer:
- DeepSWE: Flash 67.9, Pro 71.9
- AutomationBench: Flash 52.3, Pro 53.1
- JobBench: Flash 61.2, Pro 62.0
- CyberGym: Flash 95.1, Pro 94.0 (Flash wins here)
For a coding assistant or document pipeline burning hundreds of thousands of tokens per job, Flash at $0.28 out versus Opus 5.5 at $20 out is a straightforward calculation.
Where it loses and what Xiaomi published anyway
Xiaomi’s own tables show the gaps. Claude Opus 5 beats Pro on DeepSWE and ProgramBench. On Terminal Bench 4.0 the margin is wide: Xiaomi at 34.9 against GPT-6 Astra at 59.6. So the claim that Pro performs on par with Opus 5 across most agent benchmarks is Xiaomi’s own, from its own testing. The Artificial Analysis score is independent.
But what Xiaomi published alongside the weights may matter more than any ranking. In addition to the models came a full technical report, the complete RL framework, over 7,000 task environments with automatic graders, composable mini-harnesses, and a smaller distilled model. Tasks cover software development, cybersecurity, office work, web design, and roughly a thousand music composition problems. Reward design, hyperparameters, data mixtures, and training costs are all included. The stated goal is reproducibility. A published RL framework that smaller labs can run on smaller models changes what a well-funded research team can attempt. Leaderboard positions last weeks. Infrastructure like this lasts longer.
One open question: Anthropic named Xiaomi among seven Chinese labs accused of routing requests through Claude to harvest its outputs. Xiaomi has not responded to that allegation. What it has done is hand anyone who wants to verify its claims the weights, the code, the environments, and the bill.
