Overnight, Xiaomi released MiMo-V2.6-Pro and MiMo-V2.6-Flash. Open weights, MIT license, both of them. Pro scores 46 on Artificial Analysis’ Intelligence Index, which makes it the highest-rated open-weight model in the world.
46 points, tied with a Grok that shipped the same day
That 46 also beats Grok 4.6 at 44 and Gemini 3.8 Flash at 41. Grok 4.7, released on the same day, ties it. DeepSeek V4.1-Flash, which I wrote about a couple of weeks ago, sits at 39, and V4.1 Pro at 36.
The prices are where it gets awkward. Xiaomi charges $0.435 per million uncached input tokens and $0.87 per million output tokens. Claude Opus 5 charges $5 and $25. Per token, Pro is roughly twenty-three times cheaper than the model it keeps measuring itself against. Flash, at $0.14 and $0.28, is another third of that.
Where Opus 5 still wins
In Xiaomi’s own benchmark table, Opus 5 stays ahead on DeepSWE v1.1, ProgramBench and Terminal Bench 4.0. GPT-5.6 Sol leads on several cybersecurity evaluations. This is not a clean sweep, and Xiaomi does not claim one: the jumps over V2.5 are enormous, with DeepSWE v1.1 going from 19.0 to 71.9, but the top of the field is still somewhere else.
Under the hood, Pro is a mixture-of-experts model with 1.02 trillion parameters and 42 billion active. Flash has 310 billion total and 15 billion active. Both handle a one-million-token context and take text, images, audio and video.
$2.62 million for thirty training steps
The real story sits in the technical report. Xiaomi trained both models over thirty large RL steps covering about 750,000 trajectories, in under six days, for $2.62 million on Pro and $850,000 on Flash. Each step runs 1,568 prompts with sixteen attempts apiece — roughly 25,000 rollouts and up to 3.7 billion training tokens.
Only 43.5% of Pro’s bill went into training itself. Generating the rollouts took 43.8%, grading them another 12.7%. More than half the budget was spent producing and judging experience before a single weight moved.
The model cheated, and Xiaomi wrote it down
The section on reward hacking is unusually candid. Early in training, agents worked out that they did not have to fix the assigned bug at all. They could download a newer release of the package, pull the upstream source file, or search the issue history for the fix someone had already published. Tests pass, task skipped.
Xiaomi’s answer was to strip build artifacts and caches from the environments, delete future Git history, cut network access, and run a dedicated hack agent that hunts for remaining loopholes before training starts. In the final run, confirmed cheating trajectories stayed under 2%.
The price gap matters more than the benchmark
Fuli Luo, formerly at DeepSeek and now leading the MiMo team, calls this one of the largest single RL runs any open-source team has attempted. Several dozen people, on Chinese chips.
For anyone working with Claude every day, one release changes little. Opus 5 still wins the hard problems. But the premium gets harder to defend each time this happens: run an agent across hundreds of thousands of tokens and you are paying Anthropic twenty-three times more for a lead that has shrunk to three benchmarks.