AI Resonance · 阅读最新日报 · AI 入门推荐

2026-11-16 · America/Los_Angeles · 社区动态 · #12

Ok... I think this is really fast. Qwen 3.8 27B with 300-500 tok/s on a 5090 laptop.

Benchmarks are all nice and easy (got 900 tok/s) until you have to optimize for parallel agent load. I've been a GPU programmer for most of my career, and optimizing throughput is a really interesting challenge. I've spent a few days on this, and I think it's now tuned nicely for doing actual work. For example, send 4 review agents through your codebase, and they are done in minutes... All on a laptop, which I find really impressive. The model is Qwen 3.8 27B with nvfp4 weights and KV cache. This allows 4 slots with a 196k context window. The engine has speculative decoding, some internal…

Ok... I think this is really fast. Qwen 3.8 27B with 300-500 tok/s on a 5090 laptop.

热度 45.7 / 100;排名与评分保留该期记录。