Xiaomi opens MiMo-V2.6 and closes in on proprietary leaders
September 22, 2026
Xiaomi releases MiMo-V2.6 Pro and Flash with open weights under an MIT license. Strong agent benchmarks meet unresolved questions about training data and vendor claims.
What this is about
Xiaomi released the MiMo-V2.6 model family with open weights on September 22, 2026. It includes an omnimodal Pro model and a Flash variant with 309 billion parameters, 15 billion of which Xiaomi says are active per pass. The models use the MIT license and support context windows of up to 256,000 tokens.
The release is notable because a company best known for smartphones is now competing directly in high-performance open models. Independent reports place MiMo-V2.6 Pro near the top of current open-weight rankings, although several figures come from Xiaomi's own evaluation setup.
What MiMo-V2.6 actually does
MiMo-V2.6 handles text and visual inputs and targets coding, tool use and multi-step agent tasks. The Flash model uses a mixture-of-experts architecture: only a selected part of the network processes each token. That reduces compute compared with a dense model of the same total size.
Xiaomi reports 71.9 on DeepSWE v1.1, 53.1 on AutomationBench and 89.9 on Terminal Bench 2.1 for Pro. Claude Opus 5 and GPT-5.6 Sol remain ahead on some tests in the same comparison. The release is therefore not a clean benchmark sweep, but a serious open competitor with different strengths.
Why it matters
Open weights let companies and research teams run, inspect and adapt a model on their own infrastructure. The MIT license is substantially more permissive than many recent research or community licenses. Developers can reduce dependence on a single cloud API, while still facing significant hardware, operational and security requirements.
The release context is contested. Anthropic has accused several Chinese labs of using Claude outputs for large-scale distillation. Xiaomi emphasizes extensive reinforcement learning in its public material. The disclosed information does not make it possible to fully verify which data entered pretraining and post-training.
In plain language
An open-weight model is like buying a kitchen appliance instead of using a delivery service. You can run, open and modify it in your own kitchen. You still need electricity, maintenance and good ingredients, and a performance claim does not replace testing your own recipe.
A practical example
A software company wants to summarize 5,000 internal support cases per week without sending texts to an external API. It tests MiMo-V2.6 Flash on its own server against two smaller models. The team measures answer quality, hallucination rate, latency and total cost across 500 representative cases. Deployment begins only after privacy review and error rates meet the target.
Scope and limits
- Many performance figures come from the vendor and are not directly comparable with every independent setup.
- Open weights do not mean fully open training data or automatically traceable provenance.
- A model of this size requires substantial hardware and a dedicated security architecture; the MIT license does not solve operational problems.
A leading benchmark rank does not guarantee strong German output, company-specific performance or safe sensitive decisions. Reproducible tests with real, lawfully usable data remain decisive.
SEO & GEO keywords
Xiaomi MiMo-V2.6, MiMo Pro, MiMo Flash, open weights, MIT license, mixture of experts, multimodal AI, coding model, local AI, reinforcement learning
💡 In plain English
Xiaomi is releasing a capable multimodal model with open weights and a permissive license. Its results are interesting but require independent and use-case-specific testing.
Key Takeaways
- →MiMo-V2.6 Pro and Flash ship with open weights under the MIT license.
- →Flash has 309 billion total parameters and activates 15 billion per pass, according to Xiaomi.
- →Xiaomi reports strong coding and agent results but not a win across every benchmark.
- →Training data, independent reproducibility and hardware needs remain key limits.
FAQ
Is MiMo-V2.6 open source?
The weights use the MIT license. That does not automatically mean the complete training data and every training step are public.
Can MiMo-V2.6 run on a normal laptop?
The large model is too demanding for typical laptops. Quantized or hosted variants may improve access but change performance and control.
Is it better than proprietary frontier models?
It is competitive on individual benchmarks. Proprietary models lead on others, and vendor results require independent verification.