Running an Opus-level coding agent locally at 2x speed for free: a 15-page report
julianharris · x · 2026-09-05
Julian Harris published a hand-written, 15-page semi-technical report on the state of the art in local AI coding, covering two setups: a Mac M5 Max 128GB and an Ubuntu machine with an RTX 4090 24GB.
- Headline claim: he runs an Opus-level coding agent locally at nearly 2x Claude Opus speed, for free.
- On the Mac, his model pick is MTPLX.
- The report is aimed at the Local AI community and reads as a full hardware + software selection + hands-on experience writeup.
More from coding & agent
- Two years after Devin launched: Cognition's task success climbed from ~30% to 90% — silasalberti · 2026-09-05
- I built a useless Gemini chatbot, then added a "teaching layer" of lessons — now it works — Realistic-Middle7168 · 2026-09-05
- A Multi-Model "Plan-Implement-Review" Workflow Wired Through Claude Code — dr_cintas · 2026-09-05
- Viral dev joke: AI has solved writing code, not meetings, requirements or DNS — StewartalsopIII · 2026-09-05
- Git worktree workflow for running multiple branches and coding agents in parallel — JeremyCMorgan · 2026-09-05
- One dev's agent idea-exploration workspace: git worktrees plus append-only memory files — viksit · 2026-09-05