Running 35B Models at 80 tok/s on a P40

otacon6531 · reddit · 2026-07-16

The author shares benchmarks of running Qwen3.6:35B UD Q4KM on a single NVIDIA P40 24GB, reportedly achieving a burst speed of 80 tok/s while supporting a 100k context.

Key configurations include:

The author's goal is to maximize the performance of this older GPU without degrading the model's capabilities.

Original post →

More from Infra

Infra channel →