A car company just published the strongest open weights on the board
Xiaomi released and fully open-sourced MiMo-V2.6 on September 21, and the flagship Pro model lands at 46 on the Artificial Analysis Intelligence Index. That score puts it ahead of every other open-weight model on that leaderboard, above xAI’s Grok 4.6 and Google’s Gemini 3.8 Flash, and level with Grok 4.7, which shipped the same day. Xiaomi itself concedes the model still trails the strongest closed systems, naming Claude Fable 5.1 and GPT-6 Astra. The company published MIT-licensed weights for both MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL on Hugging Face, alongside a technical report.
That last part is the part worth sitting with. A year ago, a top-of-board open release meant a vendor crying benchmark victory while keeping the training details vague. Xiaomi published the training run itself, including costs, and the numbers are not what most people assume frontier RL costs.
The RL bill came to about $3.47 million
Both models finished a reinforcement learning run in under six days: thirty steps each, roughly 750,000 trajectories total. Flash cost around $850,000. Pro cost about $2.62 million. Each update step used 1,568 prompts with 16 rollouts per prompt, producing between 3.5 and 3.7 billion training tokens per step, with coding, general agent, visual, and cybersecurity tasks mixed into the same run rather than trained separately.
I keep re-reading that total. Three and a half million dollars is real money, but it is the kind of money a mid-size lab or a well-funded startup can spend, not the nine figures that frontier training folklore demands. If the accounting holds up under outside scrutiny, the effective price of a frontier-grade RL post-training run just became public record, and that changes what smaller teams can even attempt. The catch, and it is a real one: every benchmark number here is Xiaomi’s own evaluation. Independent leaderboard submissions do not exist yet, and the model card’s own tables show mixed results, with Pro trailing closed models on ProgramBench, Terminal Bench 4.0, and exploit-generation tests.
Two models, one architecture trick
Both models are sparse mixture-of-experts. Pro carries about 1.02 trillion total parameters with 42 billion active per token. Flash carries roughly 310 billion total with 15 billion active. Both take text, images, video, and audio, both claim a one million token context window, and both cap output at 128,000 tokens.
The interesting economics sit with Flash. On Xiaomi’s agent benchmarks it trails Pro by a few points, 67.9 versus 71.9 on DeepSWE v1.1, but costs about a third as much through the API: $0.14 per million uncached input tokens and $0.28 per million output, against $0.435 and $0.87 for Pro. Cache-hit input is nearly free on both. VentureBeat calls Flash the second-cheapest frontier-class model available over an API right now, and it is hard to argue with the shape of that claim given the price sheet.
The benchmark picture, with the fine print
The RL run itself moved the needle in ways that are easier to trust than a leaderboard position, because you can compare before and after on the same model. Flash improved its average pass rate on training tasks by about 25 percent in relative terms and jumped from 48.8 to about 65.7 on the held-out DeepSWE v1.1 evaluation. Pro moved from 58.4 to about 72.6 on the same benchmark. Gains on held-out tests are the honest signal that the RL generalizes rather than overfits the task pool.
Elsewhere the picture is lumpy in a useful way. CyberGym looks spectacular for Flash at 95.1, then the floor drops out on offensive-security evaluations, where Flash scores 6.0 on ExploitGym while Pro reaches 22.1, still far behind closed models. Xiaomi also released MiMo-V2.6-Distill-Qwen-9B, a 7,000-task RL environment library, and says it is open-sourcing the training environments and RL code, though the repositories shipping at launch contain weights, a technical report, and deployment instructions rather than the full training stack. There is also a hosted Pro-UltraSpeed tier, which Xiaomi claims serves output up to 20 times faster than the standard Pro endpoint at the same quality, priced well above the base model.
What to do with this
If you self-host open models, Flash is the practical one. At 310 billion total parameters it still assumes a cluster, not a workstation, but the active parameter budget is 15 billion, which keeps inference cost per token sane for a model that holds its own on agent benchmarks against systems three times its price. Teams running agent workloads on API models should also price out Flash against their current provider: a third of Pro’s cost for a few benchmark points is the kind of trade most production systems should take.
The number I would watch next is not another benchmark. It is whether anyone outside Xiaomi reproduces the DeepSWE and Terminal Bench results on the released weights, and whether the promised RL environments actually land. Self-reported numbers launched this story. Independent runs are what will finish it.