Xiaomi open-sources MiMo-V2.6 and the RL machinery behind it

Xiaomi released and open-sourced its MiMo-V2.6 series, publishing Hugging Face repositories for two omnimodal models, MiMo-V2.6-Pro and MiMo-V2.6-Flash, along with a technical report, deployment instructions, training environments and reinforcement-learning code. The Pro model card lists a sparse mixture-of-experts architecture with 1.02 trillion total parameters and 42 billion activated per token, while Flash lists 309 billion total and 15 billion activated; both accept text, images, video and audio with a claimed 1-million-token context window. Xiaomi said each model completed 30 reinforcement-learning steps across roughly 750,000 trajectories in fewer than six days, with 1,568 prompts and 16 rollouts per prompt producing 3.5 billion to 3.7 billion training tokens per step, while MiMo-V2.5 API prices were retained.

Xiaomi open-sources MiMo-V2.6 and the RL machinery behind it Xiaomi released two omnimodal model repositories, reinforcement-learning resources and a technical report while retaining MiMo-V2.5 API prices. By RuntimeWire Staff https://runtimewire.com/author/runtimewire-staff · Published · Updated Primary source: Xiaomi MiMo https://mimo.xiaomi.com/mimo-v2-6 Why it matters Xiaomi is pairing downloadable model repositories with detailed reinforcement-learning disclosures and hosted inference. Competing AI labs now face another large open model whose training claims and weights can be tested rather than taken solely from a benchmark chart. Lei Jun https://ir.mi.com/board-member-management/lei-jun?ref=runtimewire 's Xiaomi https://xiaomi.com/?ref=runtimewire has released and open-sourced its MiMo-V2.6 series https://mimo.xiaomi.com/mimo-v2-6?ref=runtimewire , publishing repositories for two omnimodal AI models trained through reinforcement learning. Lei founded Xiaomi in 2010 after founding online retailer Joyo.com, which Amazon acquired in 2004, and holding senior roles at Kingsoft. MiMo is an internal Xiaomi AI initiative that extends Xiaomi's work from phones, connected devices and electric vehicles into foundation models. The technical work is led by Luo Fuli, the former DeepSeek researcher who joined Xiaomi to head the MiMo team, VnExpress reported https://e.vnexpress.net/news/tech/personalities/chinese-ai-prodigy-luo-fuli-leaves-deepseek-for-xiaomi-4964003.html?ref=runtimewire in November 2025. Luo previously worked at Alibaba's DAMO Academy and contributed to DeepSeek's model research. Public Hugging Face repositories for Pro and Flash are accessible, and Xiaomi's launch page says the company is releasing and open-sourcing the series. OpenRouter also lists the models for hosted access. Xiaomi publishes weights and RL resources The series consists of MiMo-V2.6-Pro and MiMo-V2.6-Flash. Both accept text, images, video and audio, and both have a claimed 1-million-token context window. The Pro model card https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL?ref=runtimewire describes a sparse mixture-of-experts architecture with 1.02 trillion total parameters and 42 billion activated for each token. The Flash model card https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL?ref=runtimewire lists 309 billion total parameters and 15 billion activated. Xiaomi's release package includes a technical report https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/blob/main/MiMo V2 6 technical report.pdf?ref=runtimewire , deployment instructions and details of the reinforcement-learning system. Xiaomi says it is also open-sourcing the training environments and RL code needed to examine and reproduce the work. That is a materially broader disclosure than an API launch. Model weights let developers operate the checkpoints on their own infrastructure, although models of this size remain expensive to serve. Xiaomi's deployment recipe for Pro recommends distributed inference across multiple nodes and GPUs. Xiaomi is also rolling out MiMo-V2.6-Pro-UltraSpeed, a hosted serving option that Xiaomi says can generate output at up to 20 times the speed of the standard Pro service at the same model quality. It appears in the commercial catalog as a higher-priced inference tier. Xiaomi details a six-day reinforcement-learning run MiMo-V2.6-Pro and Flash each completed 30 reinforcement-learning steps across roughly 750,000 trajectories in fewer than six days, according to Xiaomi's account of the production run https://mimo.xiaomi.com/mimo-v2-6?ref=runtimewire . Each update used 1,568 prompts with 16 rollouts per prompt https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL?ref=runtimewire , producing between 3.5 billion and 3.7 billion training tokens per step https://mimo.xiaomi.com/mimo-v2-6?ref=runtimewire . Xiaomi mixed coding, general agent, visual and cybersecurity tasks into the same run instead of training a separate model for each category. Xiaomi's MiMo team also built graders that compare successful trajectories against one another. A binary test can tell a training system whether an agent completed a task, though it cannot distinguish a concise, reliable solution from a wasteful one that happened to pass. Xiaomi's approach ranks those successful attempts and redistributes the training reward toward higher-quality paths. Xiaomi says Flash's average pass rate on training tasks increased 25% in relative terms, while Pro improved 12% https://mimo.xiaomi.com/mimo-v2-6?ref=runtimewire . On DeepSWE v1.1, a held-out software-engineering test, Xiaomi reports Flash rising from 48.8 to 65.68 and Pro moving from 58.4 to 72.57. Those results come from Xiaomi's own evaluation and have yet to be independently reproduced. The published benchmark table is mixed enough to be useful. MiMo-V2.6-Pro trails several closed models on ProgramBench, Terminal Bench 4.0 and exploit-generation tests. It performs more competitively on automation, visual coding and some long-horizon agent evaluations. Xiaomi says Pro scored 46.32 on version 4.3 of the Artificial Analysis Intelligence Index, placing it ahead of other open models included in that comparison. Benchmark leadership can move with a new test suite, grading change or competing release. Xiaomi's stronger evidence is the combination of accessible model repositories, architecture details and reported training methods. Developers can examine the models rather than accepting a leaderboard graphic as the entire product. Xiaomi extends MiMo distribution beyond its own products Xiaomi says the V2.6 series retains the standard API prices used for MiMo-V2.5 https://mimo.xiaomi.com/mimo-v2-6?ref=runtimewire . The model cards also list availability through Xiaomi AI Studio, MiMo Desktop and MiMo Code. OpenRouter lists Pro and Flash https://openrouter.ai/xiaomi?ref=runtimewire with a 1.05-million-token context allowance and provides access through its unified API. The release follows Xiaomi's June 29th publication of MiMo-V2.5 under the MIT license https://mimo.mi.com/docs/en-US/news/latest/v2.5-open-sourced?ref=runtimewire . In that announcement, Xiaomi said it published the earlier model weights and worked with chip manufacturers, inference frameworks and agent-framework vendors on adaptation and distribution. Xiaomi's MiMo team is betting that reinforcement learning on tasks with machine-checkable outcomes can push open models forward faster than another round of general instruction tuning. Xiaomi is also distributing the models through its own software products and pricing the standard services at V2.5 levels. MiMo-V2.6 still leaves developers with practical work. Serving the open weights requires substantial hardware, Xiaomi's benchmark claims need outside replication, and UltraSpeed's performance claim depends on the hosted implementation. Engineers now have model repositories, model cards and a technical report they can inspect.