Xiaomi released MiMo-V2.6 with the training log attached — restarts, OOMs and all
Xiaomi open-sourced MiMo-V2.6 and streamed six days of its RL run: 30 steps, $3.5M, and a public log that carries the OOMs. We test the transparency claim.
Vlad Makarovreviewed and published
Xiaomi released and open-sourced the MiMo-V2.6 series on September 21 and 22, three omnimodal models the company describes as "frontier intelligence, all the modalities, built in public." The weights are out and the prices are low, and neither is the interesting part. What separates this release is the RL dashboard Xiaomi ran for six days, streaming a production training run's logs, cost counter and failure notices while the meters were still running.
What shipped, and what it costs
The family is three models: MiMo-V2.6-Pro, MiMo-V2.6-Flash and MiMo-V2.6-Pro-UltraSpeed, the last advertised as "up to 20x faster output speed at the same quality," with no independent measurement behind it. All three are natively omnimodal — repository tags list vision-language, audio and video-understanding — and Artificial Analysis describes the flagship as "an MoE model with 1.02T total parameters and 42B active parameters."
Weights are in the XiaomiMiMo organisation on Hugging Face: MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL under MIT licence tags, plus MiMo-V2.6-Distill-Qwen-9B, a 9.4B fine-tune of Qwen3.5-9B on MiMo-generated data. Community GGUF and 4-bit MLX conversions appeared within hours; download counters still read zero at capture, which says the repositories are new, not how many people took them.
Prices are unchanged from V2.5: per million tokens, Flash at $0.0028 cache-hit input, $0.14 cache-miss and $0.28 output, Pro at $0.0036, $0.435 and $0.87, UltraSpeed at $0.036, $4.35 and $8.7. Access runs through MiMo Coder, MiMo Desktop, the MiMo API platform and OpenRouter.
Six days on the counter
The dashboard is where the release stops being ordinary. It streamed two large RL runs from the trainer's logs until they stopped:
| Run | Started (UTC) | Steps | Cost reported | Tokens |
|---|---|---|---|---|
| mimo-v2.6-pro | 2026-09-15 10:32 | 30 | $2,620,670 | ~3.43B per step, ~75B total |
| mimo-v2.6-flash | 2026-09-15 15:16 | 30 | $854,044 | ~3.7B per step, ~81.4B total |
Both stopped inside the week, after roughly 750,000 trajectories each and about 753,000 samples trained, for a combined bill near $3.47 million. Xiaomi reports relative pass-rate gains of about 25% for Flash and 12% for Pro on the training tasks, and DeepSWE v1.1 up roughly 17 points for Flash, 48.8 to 65.68, and 14 for Pro, 58.4 to 72.57.
The run published its own bad days
Read the notices, not the curves. The dashboard carried lines such as: "the pro run restarted at step 17 due to a GPU OOM issue caused by expert load imbalance. we have adjusted the training parallelism strategy." Another: "there was a network connectivity issue between the pro training cluster and the grader deployment. we have restarted the run. we also removed the cyber dataset from the upcoming pro run, since we observed some bad patterns in the rollout logs." A third: the pro run "is restarting due to a vram issue on one node." The flash run was restarted from step 15 after an infrastructure error on one dataset went undetected for roughly three hours.
A job that ran out of GPU memory because some experts drew more traffic than others; a trainer and grader that could not talk across the cluster boundary; a dataset broken for three hours before anyone noticed. Closed labs make that disclosure after the fact, if at all. Here it ran live while the spend counter climbed.
The cyber rows undercut the cyber headline
The vendor's benchmark appendix is more informative than the headline. Xiaomi cites Artificial Analysis to call MiMo-V2.6-Pro, at 46.32, "the strongest open-source model to date"; Artificial Analysis' own post says "the top open weights model on the Artificial Analysis Intelligence Index (46)" at $0.13 per task. Real, but narrower than the superlative implies. The cyber rows matter more:
| Model | CyberGym | ExploitGym | ExploitBench |
|---|---|---|---|
| MiMo-V2.6-Pro | 94.0 | 17.8 | 47.9 |
| MiMo-V2.6-Flash | 95.1 | 6.0 | 25.3 |
| GPT-6 Astra | — | 42.4 | 100.0 |
| Claude Fable 5.1 | — | 30.4 | — |
| Claude Fable 5 | — | — | 78.0 |
Flash posts the column's highest CyberGym score at 95.1, above DeepSeek V4.1 Flash at 88.1 and GLM 5.3 at 84.5. On ExploitGym the same models score 17.8 and 6.0 against 42.4 for GPT-6 Astra; on ExploitBench, 47.9 and 25.3 against Astra's 100.0. Finding known weaknesses in controlled settings and building working exploits are different skills, and Xiaomi's own table separates them by a few lines.
What a public log still cannot do
The dashboard is self-reported and outsiders cannot audit the counters. The DeepSWE checkpoints were scored on Xiaomi's own setup with its own attempt counts, which makes them progress markers rather than leaderboard entries — the same benchmark under a different harness is not the same number. The restarts matter in the arithmetic too: if Pro restarted at step 17 and Flash from step 15, the 30 steps that finished are the surviving tail of a longer, more expensive job, and the discarded compute appears nowhere as a line item. Open weights are also not an open pipeline; the licence covers files, not the data recipe or the training code. The most appreciative reader on Hacker News drew that distinction, then asked for more anyway.
Reception ran ahead of the evidence
On Hacker News, the release drew roughly 271 points and 114 comments in an hour. The top comment cared about process, not capability: the realtime dashboard "was an incredible learning and teaching tool for me," rao-v wrote, praising a write-up that was "unusually comprehensive" and covered "even the stuff they didn't do well on." A reply spun the same evidence the other way, arguing this is why US labs want to slow development down; that is one commenter's reading, not a Xiaomi disclosure. Another called publishing the training spend an obvious shot at the American frontier labs. Promotional motive or not, the meter was real and public.
What would settle it
Three things would turn a good week for transparency into something durable. Independent evaluation of the final checkpoints, on harnesses Xiaomi does not control. A reproduction, by somebody else running a comparable RL setup and hitting the same classes of restart. And a second public run whose log still carries its bad news when that news is worse than an out-of-memory restart — the only version of this that costs the company something. For now the scoreboard claim belongs to the open-weights ladder and the benchmark rows feed the DeepSeek comparison; neither has been reproduced outside the lab that published it.
