Xiaomi launches MiMo-V2.6-Pro, surpassing DeepSeek as top open weights model

Xiaomi announced on Tuesday that its MiMo‑V2.6‑Pro language model has become the highest‑scoring open‑weight AI on Artificial Analysis’s Intelligence Index, posting a score of 46 and edging out DeepSeek’s previous leader. The benchmark, which aggregates results from a range of third‑party tests, placed MiMo‑V2.6‑Pro ahead of all publicly available models in categories such as reasoning, coding and multilingual understanding. The company released the model alongside a lower‑cost variant, MiMo‑V2.6‑Flash, for developers who need faster inference on modest hardware.

The breakthrough comes as the AI field shifts toward open‑weight models that can be run on‑premise or customized by enterprises. Since OpenAI’s ChatGPT and Google’s Gemini dominated the market, many firms have struggled to offer comparable performance without restrictive licensing. Xiaomi entered the arena in 2023 with its MiMo‑V2.5 series, aiming to leverage its massive hardware ecosystem and low‑cost manufacturing to undercut rivals. The latest version builds on a transformer architecture optimized for Chinese and English text, and it incorporates a new sparse‑attention mechanism that reduces compute while preserving accuracy.

Industry analysts say the result signals a maturing Chinese AI sector that can compete on global benchmarks without relying on foreign cloud services. “Xiaomi’s score of 46 puts it in the same tier as the leading closed‑source models, which is remarkable for a company better known for phones and scooters,” noted Li Wei, a senior analyst at SinoTech Research. The company also reported that early adopters of MiMo‑V2.6‑Pro have seen a 30 percent drop in latency compared with the previous version, while maintaining higher win‑rate scores on standard exams.

Xiaomi plans to make MiMo‑V2.6‑Pro available through its AI Cloud platform starting next month, with a free tier for hobbyists and a paid tier for enterprise integration. The firm says it will continue to iterate, targeting a score above 50 by the end of 2027. If the model gains traction, it could pressure other hardware manufacturers to open their own weights, accelerating a broader shift toward more accessible, high‑performing AI tools.