Xiaomi releases MiMo-V2.6 open-weight models under an MIT licence

Xiaomi published MiMo-V2.6 on Hugging Face as open weights under the MIT licence: a Pro model and a Flash model, each with a one-million-token context window. The Pro model card describes a sparse mixture-of-experts design with 1.02 trillion total parameters, 42 billion of them active, 384 routed experts of which eight are active at once, and input in text, image, video and audio. Let’s Data Science, dated 22 September, reported the Flash model at 309 billion total and 15 billion active parameters. The Hugging Face repositories were created on 21 September in UTC, which is 22 September in Beijing.

Xiaomi’s account of the training centres on reinforcement learning. The card describes one mixed run across coding, general-agent, visual and cybersecurity tasks rather than separate runs per domain, with an agentic grader that ranks passing solutions within each group of attempts and a distillation stage drawing on several teacher models. Let’s Data Science reported that the release also includes a technical report, training environments and the reinforcement-learning code.

The benchmark table is Xiaomi’s, and it is not uniformly favourable. The company reports 71.9 for the Pro model on DeepSWE v1.1, against 74.0 for Claude Opus 5, 73.0 for GPT-5.6 Sol and 70.0 for Claude Fable 5, and 82.0 on OSWorld-Verified against 83.4 for Opus 5. On Terminal-Bench 4.0 it lists 34.9 against 49.0 for Opus 5 and 39.9 for GPT-5.6 Sol. On the exploit-development test ExploitBench its 47.9 sits well below the 70.0 to 78.5 the same table gives the three US models. Xiaomi lists no comparator for its 94.0 on CyberGym. Let’s Data Science cautioned that vendor-reported benchmarks need validating against a team’s own workloads. The model card cites no independent evaluation.

The release put a consumer-electronics maker among the Chinese organisations publishing large open-weight models, alongside DeepSeek’s September release. Xiaomi’s cybersecurity scores arrived a month before Moonshot published its own cyber evaluation and data-filtering claims for an open release.