Xiaomi’s MiMo team switched on a public dashboard this week that lets anyone watch a trillion-parameter AI model train in real time — reward curves updating, token throughput climbing, running costs ticking up by the second. No frontier lab in the US has done anything like this. Anthropic, OpenAI, and Google guard their training runs like state secrets. Xiaomi just streamed one.
What You’re Looking At #
The MiMo V2.6 dashboard lives at mimo.xiaomi.com/rl and updates directly from trainer logs. Two models are training in parallel: MiMo-V2.6-Pro and MiMo-V2.6-Flash. The Pro run burns through roughly $432,000 per day; the Flash run has accumulated over $512,000 in total compute cost so far. The dashboard displays all of this publicly — step counts, reward curves, accepted sample ratios, token throughput — for anyone with a browser.
Each training step processes approximately 2 billion tokens across 1,568 prompts, with 16 fully asynchronous rollouts per prompt. The pipeline is non-blocking: rollout generation, environment execution, reward scoring, and weight updates happen independently, which is why Xiaomi can run at this throughput without hitting the synchronization bottleneck that plagues simpler RL setups.
The Number That Matters: 65.97% on DeepSWE #
MiMo-V2.6-Pro’s mid-training score on DeepSWE v1.1 — the most demanding agentic coding benchmark — is already 65.97%. Its predecessor, MiMo-V2.5, started from a 19% baseline on the same benchmark. That is a 47-point jump, and the run is not finished.
To calibrate that number: Claude Opus 4.6 sits in the mid-to-high 60s on DeepSWE variants. You are watching a model approach or match that performance while it is still in training. Whatever V2.6-Pro scores when training completes, it will be higher than 65.97%.
Why Transparent Training Is Unusual #
The standard playbook for frontier labs is to train in secret, run internal evals, and announce a finished model. Anthropic, OpenAI, and Google have all cited safety concerns as a reason for that opacity. The argument is that leaking training dynamics creates misuse risk or competitive advantage for adversaries.
Xiaomi’s counter-argument is implicit in the dashboard: accountability through visibility. If you can watch the reward curve, you can catch reward hacking. If you can see the cost, you can judge whether the compute investment makes sense. This approach aligns with what other Chinese labs — Kimi K3, Qwen 3.8, DeepSeek — have been doing in 2026: making transparency a competitive signal rather than a liability.
Luo Fuli, the team lead who joined from DeepSeek in late 2025, publicly announced the live stream on September 16. She was a core developer on DeepSeek-V2 and one of China’s most prominent young AI researchers. The fact that she is comfortable running a public RL training session at this scale says something about how Xiaomi is positioning the MiMo project.
What V2.5 Already Does (The Baseline) #
If MiMo V2.6 is the thing you are waiting for, V2.5-Pro is the thing you should be using now. The predecessor already outperformed Claude Code on SWE-bench Pro and Terminal Bench 2. It runs an OpenAI-compatible and Anthropic-compatible API, which means dropping it into an existing agent stack requires changing one URL and one API key. The pricing differential is hard to ignore: V2.5-Pro lists at $0.42 per million input tokens and $0.83 per million output tokens. Claude Opus 4.6 lists at $15 per million input tokens. Cache-hit input on MiMo drops to $0.0035 per million — a 99% reduction. For agent workloads that hit the same context repeatedly, the cost gap approaches 99.9%.
The Forkast analysis of the V2.6 training run notes that MiMo is explicitly optimized for multi-task agent environments — code, general reasoning, vision, cybersecurity, and chat mixed in the same training run. That breadth matters for production agent systems that handle more than one task category.
What to Do Right Now #
Watch the live MiMo V2.6 dashboard. Not as a novelty — as due diligence. If V2.6-Pro holds its mid-training trajectory, it will ship as a serious alternative to Claude Opus at a fraction of the price, under an MIT license that lets you run it locally. The weights will land with no usage restrictions, which matters for any workload involving sensitive code, proprietary data, or regulated industries.
Start integrating V2.5-Pro into your stack now. The migration path from V2.5 to V2.6 will be minimal — same API endpoints, same license model. Teams that wait until V2.6 ships will spend time on integration that early adopters have already resolved.
The Hacker News thread hit the front page on September 17 with hundreds of upvotes and a developer already running MiMo in production noting: “The cost is unbelievably low, and the quality of intelligence I get is equivalent to when I was working mostly with Anthropic models.” MiMo V2.6 will only widen that advantage.