Xiaomi is training MiMo-V2.6 in public — $30K an hour
Xiaomi's MiMo team has done something no Western lab has: opened the training run itself, not just the weights. The dashboard updates from the trainer's own logs.
Xiaomi is streaming the reinforcement-learning run behind MiMo-V2.6 Flash and Pro on a public dashboard — reward curves, step counts, batch composition, GPU node failures, and a running compute bill that works out to roughly $30,000 an hour. The page went live this week after MiMo lead Luo Fuli broke a six-month silence since the team open-sourced MiMo-v2.5 in April, and the models are only about a day into training: Flash started from a lower base and is closing on Pro, while Pro still holds a small lead because its steps take longer. That the numbers are visible at all is the news. Labs publish open weights, sometimes training code or data; a live RL run with the failure log attached is new, and it puts the industry's real cost curve on a public URL.
The mechanics are worth understanding because they explain the bill. Each round rolls out 1,568 prompts across 16 attempts, so a single training step can produce more than 25,000 trajectories — and an agent trajectory is not one question and one answer but dozens of turns of tool calls, environment feedback and retries, which pushes a step into the billions of tokens. Xiaomi runs those stages asynchronously: some agents are still working, others are waiting on scoring, others are already feeding the trainer. It is multi-task agentic RL, mixing code, general, vision and chat tasks into one run under different harnesses, with three scaling axes — more trajectories, more task environments, and more compute spent judging which trajectories were actually good.
Our read: the interesting claim is the scoring dimension. A code agent that runs a test suite generates its own reward signal; an agent that spent 40 steps searching, writing and fixing has no such oracle, so most of Xiaomi's compute goes into turning a coarse pass/fail into a per-step judgment. That is a bet that reward models, not raw rollouts, are the current bottleneck.
This is the second lab to make open training a spectacle — we covered the first in August, Marin opens a 535B-model training run in live public view. Xiaomi adds the cost meter and the hardware-failure ticker, which is the more honest version: a run this size is mostly a story about flaky nodes.
Arcee AI raised a Series B that values the American open-weight lab above $1 billion, led by Vista Equity Partners, Cambium Capital and Emergence Capital, with Hitachi, IAG, M12, Wipro and others participating. Arcee did not disclose the amount. The pitch is capital efficiency: the company says its entire 2025 model lineup, including Trinity Large, a 400-billion-parameter mixture-of-experts model, was built for about $20 million, salaries and infrastructure included, and that the round funds the next Trinity generation plus expanded work with the U.S. Department of Energy and its national labs. "The U.S. is far ahead in closed-source, but it kind of dropped the ball on open-source," co-founder and chief executive Mark McQuade told Fortune, arguing enterprises should not have to trade frontier capability for control of the model.
Our read: Trinity Large exists because Meta stopped shipping Llama in early 2025, and a $1 billion valuation on a $20 million training budget is a bet on that gap being a market, not a niche. The counterweight is the competition — Chinese labs are still shipping stronger open weights more often, which is the dynamic we traced in Deep Dive — Nvidia's $6 billion plan to beat China at open weights. Arcee's edge, if it has one, is enterprise-grade tooling around models buyers can host themselves.
What to watch: whether Xiaomi's dashboard keeps publishing through a mid-run restart — the moment it stops being a launch gimmick.
Would you trust a lab more for showing its training run live, or less for showing you the price tag? Tell us in the comments.
Sources: QbitAI · MiMo-V2.6 RL dashboard · LocalLLaMA discussion · Arcee AI · SiliconANGLE · Fortune
