Gizmochina reports that Xiaomi has expanded its AI portfolio across language, reasoning, code, and voice technologies. The article says Xiaomi first entered the open-source LLM race in April 2025 with MiMo, a 7 billion-parameter model; Gizmochina reports the reinforcement-learning variant scored 95.8% on the MATH-500 benchmark and reportedly outperformed OpenAI's o1-mini and Alibaba's Qwen-32B-Preview on AIME 2024 and 2025. Gizmochina also reports the company released a larger MoE model, MiMo-V2-Flash, in December 2025, described as a 309 billion-parameter model with only 15 billion parameters active at inference, rivaling top-tier models on software-engineering tests while achieving 150 tokens per second and an inference cost reportedly at 2.5% of Claude's, with API input pricing reported at $0.1 per million input tokens. Gizmochina documents training corpus sizes, MIT licensing, Hugging Face availability, and credits Luo Fuli as development lead. The roundup also covers miclaw, a system-level autonomous agent built on MiMo-V2-Pro, and OmniVoice, an open-source multilingual voice cloning model supporting over 600 languages.
What happened
Gizmochina published a roundup documenting Xiaomi's full AI model portfolio as of June 2026. The article reports Xiaomi entered the open-source LLM race in April 2025 with MiMo, a 7 billion-parameter model released under an MIT license on Hugging Face. Gizmochina reports the reinforcement-learning variant of MiMo-7B scored 95.8% on the MATH-500 benchmark, reportedly outperforming OpenAI's o1-mini and Alibaba's Qwen-32B-Preview on AIME 2024 and 2025. The article credits Luo Fuli as development lead and reports training figures of 200 billion reasoning tokens across a 25 trillion-token corpus.
MiMo-V2-Flash
Gizmochina reports that MiMo-V2-Flash, released in December 2025, is a sparse Mixture-of-Experts model with 309 billion total parameters and roughly 15 billion active at inference -- a design that keeps per-inference compute low while maintaining large model capacity. Gizmochina reports the model ranked among the top two open-source models on reasoning benchmarks, rivaled Claude Sonnet 4.5 on SWE-Bench Verified software-engineering tests, achieved around 150 tokens per second, and priced API input at $0.1 per million tokens, described as roughly 2.5% of Claude's API cost. Weights and code are available on GitHub and Hugging Face under MIT license; an arXiv technical report (2601.02780) covers the architecture in detail.
miclaw and OmniVoice: The roundup also covers miclaw, a system-level autonomous AI assistant announced in March 2026 that embeds MiMo-V2-Pro into the device OS to control apps, navigate mobile browsers, and manage IoT devices. OmniVoice, open-sourced in May 2026, is a multilingual voice cloning model supporting over 600 languages with emotional transitions and regional dialect synthesis.
Benchmark verification
The cited performance and cost figures are Xiaomi-reported. Independent replication of the SWE-Bench Verified and AIME 2024/2025 results will be needed to confirm Xiaomi's claims.
Scoring Rationale #
Single-source Gizmochina roundup summarising Xiaomi's LLM portfolio covering releases from April 2025 to May 2026. The underlying models were notable at release -- MiMo-V2-Flash's claimed cost-performance ratio rivaling Claude Sonnet 4.5 on SWE-Bench is significant if independently confirmed -- but this article is a retrospective explainer, not a new announcement. Solid-tier as a useful practitioner reference on an expanding open-weight model family.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.