AI BriefingXiaomi Releases Real-Time Post-Training Dashboard for Mimo 2.6; Pro Model Scores 65.78 on DeepSWE v1.1
Key point
Xiaomi has released a real-time post-training dashboard for Mimo 2.6, with the Pro model recording a score of 65.78 on DeepSWE v1.1.
Details
Xiaomi has released a dashboard that allows monitoring of the real-time post-training progress for the Mimo 2.6 model. This dashboard reflects metrics by training stage, processing status by dataset, and hardware-related issues in real time.
Benchmarks and Training Status
In the latest results based on DeepSWE v1.1 (mini-swe-agent, avg@3), mimo-v2.6-pro recorded a score of 65.78, and mimo-v2.6-flash recorded 60.77. Both models are currently undergoing training, with the Pro model in the rollout stage of step 15 and the Flash model in the training stage of step 19.
Dynamic Sampler and Data Processing
According to the Dynamic Sampler logs, the pass rate for the Flash model at step 19 is maintaining approximately 0.600. Overall, the accepted data has exceeded the target (1,568), with major datasets such as code, visual, and general meeting their quotas. However, there are also records of training restarts due to VRAM issues on some nodes or infrastructure errors.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.