Xiaomi opens MiMo-V2.6 weights and the RL machinery behind its frontier-scale training run

Xiaomi released the MiMo-V2.6 family on September 22, including MiMo-V2.6-Pro and MiMo-V2.6-Flash, and published model weights plus reinforcement-learning code and task environments. Pro is a sparse multimodal mixture-of-experts model with roughly 1.02 trillion total parameters and 42 billion active per token; Flash is substantially smaller. The release follows Xiaomi’s public post-training run, which exposed step-level progress and operational failures. Xiaomi reports that the Pro and Flash RL runs cost about $2.62 million and $854,000 respectively. Artificial Analysis independently places Pro at the top of the current open-weight Intelligence Index cohort with a score around 46, although Xiaomi’s task-specific benchmark claims remain vendor-produced.

Open weights are now common enough that the differentiator is increasingly what else a lab lets builders inspect. MiMo-V2.6 combines a frontier-scale permissive checkpoint with unusually visible reinforcement-learning infrastructure and economics. That gives researchers and smaller labs concrete material for studying agentic post-training, reward environments and training-system failures rather than only reverse-engineering a finished model. For application builders, the release also pushes capable multimodal and coding-oriented inference further down the cost curve, but the largest checkpoint still demands serious serving hardware and Xiaomi’s own benchmark tables should not be treated as independent proof.

Xiaomi’s release includes model weights, reinforcement-learning code, task environments and technical material describing the training setup. That makes the post-training recipe inspectable in ways that ordinary open-weight launches are not. The useful comparison is with projects such as K2 Horizon that are expanding 'open model' from downloadable inference weights toward reproducible training artifacts.

The public RL run recorded operational problems rather than presenting only a clean final curve. Published run notes include a Pro restart after a GPU out-of-memory event caused by expert load imbalance, a network connectivity failure between the training cluster and grader deployment, and removal of a cyber dataset after undesirable rollout patterns. That kind of operational evidence is unusually useful to teams building large-scale RL systems because it exposes failure modes normally omitted from model cards.

Xiaomi reports roughly $2.62 million of RL compute for Pro and about $854,000 for Flash, for a combined post-training bill near $3.47 million. Those figures are vendor-reported rather than audited, but they put an order of magnitude around what a large agentic RL run can cost and provide a more concrete input for model-development economics than an unspecified training budget.

Artificial Analysis places MiMo-V2.6-Pro around 46 on its current Intelligence Index, at or near the top of open-weight models. That does not validate every Xiaomi benchmark claim, and benchmark position can move quickly, but it provides external evidence that the released weights are competitive rather than merely a transparency demonstration.