The fact
The model matches Qwen2.5's base performance but lags behind after fine-tuning.
Developed through a collaboration between Renmin University researchers and ByteDance, it explores an alternative to dominant autoregressive architectures.
Click the link to read an article on the topic: