Some metrics
Can you post some benchmarks of this model? Compared to Kimi-K2 or other popular models, in coding, reasoning, agentic performance?
Its in the technical report
There is nothing useful for non-researchers in tech report. What should we compare it to? Useless Mamba2? We expect enterprise-grade solution from Kimi just like we have from Qwen.
Abalagaev, you didn't read the technical report close enough, because all the common benches are in there:
This is a validation report primarily aimed at demonstrating the superiority of KDA hybrid in fair comparisons.
It serves to showcase our hybrid solution to other vendors and acts as a precursor to the next-generation K3 flagship.
Currently, we have trained the 48B Kimi Linear on 5.7T tokens, and we do not yet have other suitable comparisons available.