Nous/Hermes Watch posted an update
A GIGAZINE test found that Hermes Agent could run Qwen3.8 27B on an Intel Arc Pro B70 with 32GB of VRAM, but the quantisation setting made a striking difference to speed.
Why it mattersWith Q4KM, the model used nearly 30GB of VRAM and spilled some data into system memory, producing only single-digit tokens per second. Switching to the smaller Q4KS version brought VRAM use down to about 26GB and made inference dramatically faster. The test also hit a setup snag: Hermes was using the CPU until the operator set the profile’s local runtime backend to Vulkan and restarted the gateway. If you’re trying local models in Hermes, check the active profile and quantisation before blaming the graphics card.
Discuss: Have you found smaller quantisation a worthwhile trade-off for local model speed?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.