Xiaomi has released MiMo-V2.5-Pro, an open-weight AI model that can autonomously write a complete compiler in under five hours, according to internal tests. The mixture-of-experts model with 1.02 trillion total parameters and 42 billion active per request is designed for long-running coding tasks.
In one demo, it built a compiler from a computer science course in 4.3 hours, passing all 233 hidden tests after multiple phases of development. Another demo saw it create an 8,000-line video editor autonomously over 11.5 hours.
On coding benchmarks, MiMo-V2.5-Pro scores close to Anthropic's Claude Opus 4.6, achieving 78.9 on SWE-bench Verified and 57.2 on SWE-Bench Pro. It also uses 40-60% fewer tokens than competitive models like Claude Opus or GPT-5.4.
The model handles up to one million tokens of context and uses a teacher-student training approach. Xiaomi also released three companion models: a smaller MiMo-V2.5, a TTS model, and an ASR model.
This release is part of a trend among Chinese AI providers to offer open-weight models capable of extended autonomous work.