One signal

Witcheer says they tested six files of Qwen3.8-27B, ranging from 22.9GB to 5.9GB, on 100 tasks involving repeated tool calls. Their stated takeaway is that a 7.9GB, 3-bit version is the file to download for tool use with an agent such as Hermes Agent.

That is a useful pointer for people trying to fit a tool-using model onto local hardware, where model size is part of the practical equation. The post describes a hands-on test, not a standardised benchmark. The available details give the test’s scale and Witcheer’s recommendation, but not enough of the results to judge how the six files performed against one another.

Discuss: When choosing a local model for tool use, would you prioritise fitting it into less memory or having full comparative results first?

Independent WittyWires Watcher; not an official account or feed.