Xiaomi open-sources MiMo-V2.6 in a 1M-token push

Xiaomi's MiMo team has published the MiMo-V2.6 Pro and Flash checkpoints as open weights, putting both models on Hugging Face on 21 September under an MIT licence. They arrive with a technical report, Scaling Reinforcement Learning Toward Self-Improvement, and a third, much smaller model distilled from the same work.

What Xiaomi has released

Three checkpoints are live in the XiaomiMiMo organisation: MiMo-V2.6-Pro-RL as the flagship, MiMo-V2.6-Flash-RL as the efficiency-balanced sibling, and MiMo-V2.6-Distill-Qwen-9B, a 9B distillation built on Qwen's architecture. None of them is gated or private, all three carry MIT terms, and the weights ship pre-quantised to fp8, spread across expert-parallel shards alongside a speculative draft model and a separate audio tokenizer.

The report's architecture table describes a 70-layer transformer that interleaves 60 sliding-window attention layers with 10 global attention layers, a hidden size of 6,144, a 128-token sliding window and 384 routed experts of which eight fire for any given token. Xiaomi quotes a maximum context of one million tokens. Vision runs on a MiMo ViT tower with 681 million parameters across 28 layers, and audio is handled by a 308-million parameter tokenizer with 20 residual vector quantiser codebooks plus a 127-million parameter patch encoder.

The model is pitched as omnimodal: text, images, video and audio through one set of weights, with the repository listing long-context, vision-language and video-understanding tags. Xiaomi also lists a speculative decoder that predicts seven further tokens per forward pass so they can be verified in parallel.

One reinforcement-learning run, four domains

The headline claim is about how the model was trained rather than how big it is. Rather than running separate reinforcement-learning passes for coding, agents, visual work and security, MiMo-V2.6 used a single mixed run with tasks and agent harnesses sharing the same batch, so that strategies learned in one domain carry into harnesses the model never saw during training.

That run used asynchronous Group Relative Policy Optimization on very large batches, which the report measures at 1,568 prompts with 16 rollouts each per step and 2.7 to 3.7 billion tokens per step. Because a pass or fail signal cannot rank two solutions that both pass, Xiaomi scaled the reward itself with groupwise agentic grading: rubrics are built offline from contrasting rollouts, passing trajectories are ranked online, and advantage is pushed towards the tidier answer. The report also says the mixture-of-experts router was frozen and hardening added against reward hacking.

The numbers are Xiaomi's own

The report's benchmark table puts MiMo-V2.6 Pro at 71.9 on DeepSWE v1.1 against 67.9 for Flash, 89.9 and 87.6 on Terminal Bench 2.1, 82.0 on OSWorld-Verified, 53.1 on AutomationBench and 94.0 on CyberGym. The same table carries columns for Claude Opus 5, GPT-5.6 Sol and Claude Fable 5, and MiMo-V2.6 edges ahead of all three on CyberGym and MiMo VisualCoding while trailing them on ExploitBench and ProgramBench. Every figure is vendor-reported from a self-published technical report, so treat the comparisons as Xiaomi's claims rather than an independent result.

What it takes to run

Xiaomi points users at SGLang, publishing an example command that runs Pro across 16 tensor-parallel ranks, two data-parallel groups and 16 expert ranks using the DeepEP all-to-all backend. Both flagship repositories carry custom modelling code, so loading them means trusting the remote code flag. The Distill-Qwen-9B checkpoint is the practical entry point for anyone without a rack full of accelerators.

Our opinion

The open-weights argument has been stuck on licence terms for a while, and an MIT grant on a trillion-parameter-class family is a real concession rather than a marketing one. The more interesting admission is the training story: telling researchers that one mixed reinforcement-learning run beat four separate ones, with the environment diversity named as the reason, hands competitors the recipe as well as the result. That is the part of MiMo-V2.6 worth copying, and it has nothing to do with parameter counts.

The deployment recipe does undercut the word open, though. Sixteen expert-parallel ranks for the flagship is a data-centre configuration, so what Xiaomi has really released is a model that universities with cluster access can reproduce and everyone else can only consume through an API. The distilled 9B checkpoint is the honest measure of how much the community gains here, and the vendor's own benchmark table should not be mistaken for proof that the gap to the closed models has closed.