# Xiaomi opens MiMo-V2.6 weights and the RL machinery behind its frontier-scale training run Xiaomi has released MiMo-V2.6-Pro and Flash under permissive open-weight terms alongside reinforcement-learning code and task environments, after exposing unusually detailed logs from the post-training run. Independent benchmarking places Pro at the front of the current open-weight pack, while Xiaomi’s own training record puts the two RL runs at roughly $3.47 million combined. MiMo-V2.6 is more useful than another benchmark launch because builders get both capable multimodal weights and a rare view into the reinforcement-learning machinery that produced them: code, environments, run costs and even failure notes from the training cluster. - Status: Active - Published: 2026-09-22T21:30:31+12:00 - Updated: 2026-09-22T21:30:31+12:00 - Categories: Artificial Intelligence, AI Models, AI Coding, Open Models, Inference & APIs - Tags: Coding agents, MiMo-V2.6, open weights, reinforcement learning, Xiaomi - Canonical HTML: https://beyondthe.news/dossiers/xiaomi-mimo-v2-6-open-weights-rl-training-code-costs ## What changed 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. ## Why it matters 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. ## The release exposes more than the final checkpoint 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 training log includes the ugly parts 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. ## The economics are visible enough to reason about 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. ## Independent ranking helps separate the release from its marketing 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. ## Key details - MiMo-V2.6-Pro and MiMo-V2.6-Flash were released September 22, 2026. - Pro is a sparse MoE with roughly 1.02T total parameters and 42B active parameters per token. - The family supports multimodal input including text, image, video and audio with text output. - Xiaomi released weights together with reinforcement-learning code and task environments. - The public RL run exposed operational events including GPU OOM/expert-load imbalance and grader-network failures. - Xiaomi reports about $2.62M of RL cost for Pro and $854K for Flash. - Artificial Analysis independently places Pro at roughly 46 on its current Intelligence Index, leading or tying the open-weight cohort. - Xiaomi says API pricing remains at the previous V2.5 level; builders should verify current regional/API pricing before production decisions. ## Builder takeaways - If you are researching agentic RL, inspect the released environments and training code rather than treating MiMo only as another inference checkpoint. - Use Xiaomi’s public run costs as an order-of-magnitude reference, not an audited cost model; hardware pricing, utilisation and grader infrastructure materially change the economics. - Reproduce coding and agent benchmarks in your own harness before routing production workloads to MiMo based on vendor tables. - For self-hosting, evaluate active-parameter memory, KV-cache requirements and serving stack support; open weights do not make a trillion-parameter MoE a small deployment. - Compare the released RL artifacts with other reproducibility-oriented projects such as K2 Horizon to see which parts of the training lifecycle are actually reproducible. ## What to watch - Independent reproduction of Xiaomi’s coding, agentic and multimodal benchmark claims. - Whether outside researchers can reproduce useful RL improvements using the released environments and code at smaller scale. - vLLM, SGLang, llama.cpp or other serving support and real-world throughput for the released checkpoints. - Derivative fine-tunes and distillations that make the training work practical on smaller hardware. - Whether other major model labs respond by publishing training-system logs, costs and failure notes rather than only weights and benchmark tables. ## Uncertainties - Xiaomi’s detailed RL costs and task-specific benchmark improvements are self-reported. - Artificial Analysis provides useful independent comparison but remains one benchmark methodology and does not establish superiority across all workloads. - The largest Pro checkpoint requires substantial infrastructure despite sparse activation. - The practical reproducibility of the RL pipeline depends on datasets, graders, compute and other resources beyond merely publishing code. ## Sources - [MiMo-V2.6 release](https://mimo.mi.com/docs/en-US/news/latest/v2-6) — Xiaomi MiMo · primary · 2026-09-22T00:00:00+12:00. Official release details, model capabilities, open-source artifacts and API naming. - [Xiaomi released MiMo-V2.6 with the training log attached — restarts, OOMs and all](https://traictory.com/news/2026-09-22-xiaomi-mimo-v2-6-open-source) — Traictory · specialist_independent · 2026-09-22T00:00:00+12:00. Independent inspection of the public RL logs, reported costs and operational failure notes. - [Better than DeepSeek: Xiaomi's MiMo-V2.6-Pro debuts as the top open weights model](https://venturebeat.com/technology/better-than-deepseek-xiaomis-mimo-v2-6-pro-debuts-as-the-top-open-weights-model-in-the-world-alongside-cheaper-v2-6-flash) — VentureBeat · independent_reporting · 2026-09-21T00:00:00+12:00. Independent coverage referencing Artificial Analysis ranking and release positioning.