Xiaomi releases MiMo-V2.6-Pro, an open-weights 1T-parameter omnimodal model trained via a $2.6M RL run, topping open-weights intelligence benchmarks

Xiaomi's MiMo team released MiMo-V2.6-Pro and a smaller MiMo-V2.6-Flash variant as open-weights models (1.02T total / 42B active parameters for Pro), under an MIT licence, alongside a technical report, RL training code and environments. Third-party benchmarker Artificial Analysis ranked Pro as the top-scoring open-weights model on its Intelligence Index (46), priced at $0.435/M input and $0.87/M output tokens. Xiaomi says the RL training run took 130 hours on 75B tokens at roughly $2.6M, though the article headline rounds this to $3M. The full 7k+ task RL datasets have not been released.

This gives developers and enterprises a cheaper, top-ranked open-weights alternative to evaluate against closed models like GPT and Claude, and the published pricing makes cost comparisons immediately actionable. For competitors and investors, it is another data point that a Chinese lab has matched or beaten prior open-weights leaders, adding pressure on companies competing in that tier. The bigger claim, that RL post-training is now a cheaper route to frontier-adjacent capability than pretraining scale, is plausible but rests on one vendor's self-reported figures rather than independent confirmation.

The $2.6M training cost is Xiaomi's own figure, unverified, and the headline inflates it to $3M without explanation, an early sign of imprecise reporting. The framing that RL environments are "the new pretraining corpora" and that this shows China "closing the gap" with closed frontier labs is analyst commentary layered on top of the release, not something the release itself demonstrates. The evidence arrives through a newsletter aggregator (Latent Space/AINews) with no primary source links captured, so the underlying model card, technical report and benchmark methodology have not been independently checked here.

1. 01Whether Xiaomi releases the full 7k+ task RL training datasets, not just the code and environments, which would let others actually reproduce the $2.6M cost claim
2. 02Independent benchmark runs (beyond Artificial Analysis) confirming or disputing MiMo-V2.6-Pro's Intelligence Index ranking and real-world coding/reasoning performance
3. 03Adoption signals over the next 1-3 months: developer usage, API traffic, or enterprise deployments citing MiMo-V2.6-Pro versus incumbent open-weights models like DeepSeek or Llama