DEV TOOLS Signal 185
Experiences with local models for coding
For engineers evaluating local models, RAM is the binding constraint and quality is still hit-or-miss compared to cloud-hosted models. The gap between autocomplete and agentic coding viability is substantial—smaller models may work for completion but struggle with the tool calling that agentic use requires.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
RAM is the core constraint; models in the 15–25 GB range run comfortably on 48–64 GB machines, but pushing toward 30 GB or beyond risks crashes or severe slowdowns.
Response speed has improved dramatically over the past year on Apple Silicon, though it degrades as conversation length grows.
Tool calling for agentic coding remains unreliable, with frequent failures, though models can sometimes self-recover; autocomplete is far more viable than agentic use on local hardware today.
THE CLUSTER