2026-11-02 · America/Los_Angeles · 社区动态 · #4

Not another tok/s post: local quants vs Claude on 9 real bugs

There's a lot of talk here about not chasing tok/s by lobotomizing the model, and most of what I post is still speed numbers. So this is the other half: what the models I run at home actually get done on my own work, with Claude models added as fixed reference points. Everything local runs on my own engine (capy-flash) on 2x RTX 5070 Ti 16 GB, Ryzen 9900X, 96 GB DDR5. No expert skipping for flash next. The tasks I took 9 real regressions from the private history of that engine (C++/CUDA). For each one I checked out the code right before the fix and wrote the bug report the way I'd file it to…

热度 45.4 / 100；排名与评分保留该期记录。