Xiaomi releases MiMo-v2.5 Family weights with strong coding and agent benchmarks
6 hours ago
- #AI Model
- #Coding Performance
- #Open Source
- MiMo-V2.5-Pro, an AI model from Xiaomi, completed Peking University's multi-week compiler project in Rust in 4.3 hours, passing all 233 hidden tests.
- It demonstrated strong agentic capabilities by building a video editor in 11.5 hours and designing an analog circuit in about an hour, meeting graduate-level standards.
- The model shows discipline with self-correction, such as catching and fixing a regression during compilation without human intervention.
- Benchmarks indicate competitive coding performance, closely matching models like Claude Opus and GPT-5.4 in tasks like SWE-Bench Pro and Terminal-Bench 2.0.
- MiMo-V2.5-Pro uses a hybrid attention design for efficient 1M token context handling, reducing KV cache storage by nearly 7x compared to standard attention.
- It offers token efficiency, using 40-60% fewer tokens than competitors like Claude Opus for comparable results, impacting production cost viability.
- A multimodal version, MiMo-V2.5, handles text, image, video, and audio with 1M token context, though V2.5-Pro is better for coding tasks.
- The model is open-source under an MIT license, available on HuggingFace, and targets developers needing long-horizon agentic coding workflows and cost-effective deployment.