Which local LLM should run a long-lived agent on 36GB of VRAM?

hovikyan · reddit · 2026-07-28

This is a repost of the same question about choosing a local LLM for an agent that can keep working on software projects and non-coding tasks on a machine with roughly 36GB of VRAM. The candidates are the same set of local models: Qwen3.6 27B, Qwen3.6 35B A3B, Gemma 4 31B, GLM 4.7 Flash, and Llama 3.3 70B.

Related event: Selecting Local Agent Models for 36GB VRAM(2 posts)→

Original post →

More from coding & agent

coding & agent channel →