Xiaomi MiMo-V2.5-Pro officially OSS surpasses DeepSeek-V4; Kimi K2.7-Code enters competition
Key Questions
What is Xiaomi MiMo-V2.5-Pro?
It is a 1.02T MoE model with 42B active parameters released as open-source on Hugging Face. The model leads agentic coding and SWE benchmarks while supporting a 1M context window and running efficiently on 32-64GB RTX GPUs.
Has MiMo Code been open-sourced?
Yes, MiMo Code is now fully open-sourced, enabling direct deployment and fine-tuning. This release follows its strong results in coding-focused evaluations.
How does MiMo compare to Kimi K2.6 and GLM 5.1?
Recent shootouts highlight MiMo's competitive performance, with quantization effects clearly visible. It also adds llama.cpp RPC support for multi-machine inference.
What were the findings from re-testing DeepSeek v4 Flash?
The re-test produced confusing benchmark results that differed from earlier reports. A June 2026 roundup still positions MiMo as the leading open-source model overall.
Where is the MiMo-V2.5-Pro model available?
The model weights are released on Hugging Face for community access. Users can download and run it locally or via compatible inference frameworks.
1.02T MoE (42B active) HF release leads agentic coding/SWE with 1M ctx, 32-64GB RTX optimized. New shootout vs Kimi K2.6 & GLM 5.1 shows quantization impact and llama.cpp RPC multi-machine support. A recent re-test of DeepSeek v4 Flash shows confusing results. A June 2026 roundup confirms MiMo as top open-source model. MiMo Code now open-sourced. New: Kimi K2.7-Code released with 32B active params, 30% fewer reasoning tokens, fits 32-64GB when quantized, Modified MIT license.