Xiaomi's MiMo v2.6: the AI model everyone's praising for its paperwork, not its scores
Hacker News is buzzing about Xiaomi's openness in training a new AI model — but 'transparent' isn't the same as 'independently verified', and the benchmarks are still Xiaomi's own.
Xiaomi has released MiMo v2.6, its latest open-weights language model — and the internet’s reaction says more about the state of AI reporting than about the model itself. Rather than fighting over leaderboard positions, the story dominating Hacker News (740+ points, 338 comments and counting) is that Xiaomi actually showed its working.
What Xiaomi actually did
Alongside the model release, Xiaomi published a live dashboard during the model’s reinforcement-learning training run, letting anyone watch metrics update in something close to real time. It also released a tech report going into the methodology behind the model, including, notably, benchmarks where MiMo v2.6 didn’t perform well. Commenters on the Hacker News thread compared this favourably to the norm among big US labs, where model “cards” or reports can run to hundreds of pages of dense prose with far less insight into the actual training process.
One commenter went as far as calling for a “copy as MD” button on future reports, and speculated the dashboard was likely built using verl, an existing open-source framework for reinforcement-learning training that already ships with similar visualisation tools — so this may be less of a from-scratch innovation than it first appears.
The catch: transparent isn’t the same as open
It’s worth being precise about what’s actually on offer here. As the top comment on the thread itself points out, there’s a difference between an “open” model in the fullest sense — open weights, open training data, and open training code — and what Xiaomi has done, which is releasing weights plus unusually detailed documentation and a training dashboard. That’s genuinely useful for researchers trying to learn how these systems are built, but it isn’t the same as being able to reproduce the model from scratch.
There’s also a live debate in the comments about whether Xiaomi’s headline benchmark score, on a coding-focused test called DeepSWE, is being presented fairly. One commenter noted the score curve looks like it’s flattening out, and that the reported figure matches a previous spike that later dropped — raising a mild suspicion of cherry-picking the best checkpoint rather than the final, stable one. Nobody in the thread claims to know for certain why Xiaomi stopped training when it did; guesses range from compute budget to simply not wanting to be beaten to a release by a rival lab.
Who this actually matters to
This isn’t a consumer product story. There’s no phone, app or gadget here — MiMo v2.6 is a model aimed at AI researchers, developers building on open-weight systems, and people curious about how large language models are actually trained. If you don’t work with AI tooling directly, nothing about your day-to-day tech changes because of this release.
The takeaway
Xiaomi deserves credit for showing more of its training process than most major AI labs bother to. But “more transparent than the competition” is a low bar, and the benchmark numbers are still self-reported. Anyone relying on MiMo v2.6 for real work should treat Xiaomi’s own figures as a starting point for testing, not a verified result.