3.8 wasn’t released a year ago. Of course base models change. But the quantization techniques carry over to new models. For instance, one of the improvements last week was someone porting the techniques used for macaq to qwen3.8-27b. It’s pretty satisfying to get a working implementation of a new technique barely one week after the original paper hit arxiv.
Also improvements to llama.cpp stay, and many small improvements stack up nicely.
Also improvements to llama.cpp stay, and many small improvements stack up nicely.