I have been taking different local open-weight LLMs for a test drive in different harnesses (Qwen-Code, Codex, Claude Code).

30B Mixture-of-Experts models are kind of a nice sweet spot and can solve challenging problems. And they get roughly 40 tok/sec on a Mac or DGX Spark, which is similar to GPT 5.5 in a Pro subscription and totally usable for everyday work.

More interesting is also the harness choice! Claude Code seems to be using 2x as many tokens as Codex.

Gemma 4 E2B is here just for reference to show that the tasks can’t be trivially solved by smaller models.

The longer write-up is now available at Using Local Coding Agents.

Chart from the original Substack note, comparing token use and task success across the same five local-agent tasks.

Source: lightly edited website version of my Substack note.