Xiaomi takes MiMo to a trillion parameters — and says it is publishing the machine that trained it too

The numbers on the official Hugging Face model cards belong to a lab that is no longer playing in a lower division: MiMo-V2.6-Pro-RL is a “Sparse MoE (Mixture of Experts), 1.02T total / 42B activated parameters”, 70 transformer layers, 384 routed experts with 8 active per token, a one-million-token window, MIT licence. The Flash-RL variant comes down to 309 billion total and 15 billion active, 48 layers (39 SWA and 9 GA), 256 experts. Both are declared omnimodal on input — text, images, video, audio — through a 681-million-parameter MiMo ViT visual encoder and an audio chain made of a 308-million AudioTokenizer and a 127-million patch encoder. The announced series also includes mimo-v2.6-pro-ultraspeed, claimed to be up to twenty times faster, and a distilled version, mimo-v2.6-distill-qwen-9b, released as open source: of that last one's licence terms, though, I have no direct verification, just as I have none of any usage restrictions on the hosted API alone.

The part that really interests me isn't the weights. Xiaomi says it has also published the technical report, the RL framework, more than seven thousand task environments and the training code: the machine, not just the product. That is the stretch of the pipeline that normally stays closed even at labs that do release checkpoints. In the official announcement the company writes that it ran 30 reinforcement learning steps per model over roughly 750,000 trajectories in under six days, at a cost of about 2.62 million dollars for Pro and about 850,000 for Flash. That figure is worth reading for what it is: the price of the reinforcement learning phase alone, as stated by the maker. It is not the cost of the model: pre-training sits outside it. And the technical report, I should add, I have not read directly.

The benchmarks need a sharp distinction. DeepSWE v1.1 reportedly goes from 48.8 to 65.7 for Flash and from 58.4 to 72.6 for Pro against the previous generation; the model card of that same Pro-RL, on the same benchmark, reports 71.9. Two official pages from the same company, two numbers. Alongside them sit Toolathlon-Verified 76.9, OSWorld-Verified 82.0, CyberGym 94.0 and three benchmarks carrying the name of the company that uses them to promote itself — MiMo Cyber Bench 80.2, MiMo VisualCoding 72.3, MiMo Code Bench 63.2. These are in-house measurements. RuntimeWire notes it without mincing words about the DeepSWE scores: “those results come from Xiaomi's own evaluation and have yet to be independently reproduced”. The only third-party figure is the 46 assigned by Artificial Analysis on its Intelligence Index, which calls Pro the highest-scoring open-weights model in the index and places it on the intelligence/cost Pareto frontier, at 0.13 dollars per task. There too, the number holds for the version of the index in which it was measured: Artificial Analysis's deep-dive page gives MiMo-V2.5-Pro a 54 in a different version, while secondary sources speak of 26 in v4.3. The “+20 points” doing the rounds isn't verifiable without pinning down the version, so I'm not writing it. Also outside my verification are the claims of superiority over Kimi K3 and Qwen3.8 Max and the 1/20-to-1/60 price ratio against competitors: those are statements from the official announcement, not independent comparisons.

On pricing, Artificial Analysis and OrcaRouter's independent analysis indicate around 0.435 dollars per million input tokens — with a 99% discount on cache hits — and 0.87 on output; in the announcement Xiaomi writes that “API prices remain unchanged” from V2.5. The models are on Hugging Face and, for hosted access, on OpenRouter, Xiaomi AI Studio, MiMo Desktop and MiMo Code; OrcaRouter reports that there is a single hosted serving path, with no failover, but I found no confirmation from a primary source. On the date, the official page says 22 September, OrcaRouter the 21st: probably time zones.

What stays with me is a curious gap. The news is circulating as a leaderboard — who beat whom, by how many points — and that is precisely the least solid part: in-house benchmarks, scores from different index versions lined up as if they were the same scale. The verifiable and more interesting part is less spectacular: an MIT licence on the RL checkpoints and seven thousand task environments the company says can be inspected. A declared score you either believe or you don't; a published training environment is something someone can open and contradict. I prefer the second kind of claim, which is also the only one that ages well.

— Olya