Local AI coding hits only 20-35% of Sonnet's speed in weeks-long app-build tests
julianharris · x · 2026-10-11
After weeks of testing various local AI setups (up to 128GB), the author finds local models are now smart enough to be an everyday substitute for frontier models — but far too verbose.
In long-form full app builds, no local setup exceeded 35% of the Claude (Sonnet) family's productivity, with most landing at 20-25%. A run of 11 user stories that Sonnet finished in 3 hours took some local models over a day.
The author argues optimization efforts should focus on vastly more efficient output, and notes the finding includes MLX 1.5.
More from coding & agent
- Uber details its MCP Gateway: a unified layer for hundreds of AI agent teams — bibryam · 2026-10-11
- auto-optimize speeds up genomics tool fimo ~20x with bit-identical results — jmschreiber91 · 2026-10-11
- Pointing auto-optimize at genomics code: tomtom made 7.8x faster, bit-identical output — jmschreiber91 · 2026-10-11
- Microsoft launches MXC Execution Containers to sandbox AI agents across Windows, Linux and macOS — TheKanter · 2026-10-11
- Wes Bos Open-Sources Guide to Rooting Locked-Down Android Photo Frames, With AI Agents Helping — natesiggard · 2026-10-11
- Podcast: A 4-Step Agentic Engineering Workflow — Isolate, Build, Prove, Ship — petergyang · 2026-10-11