1.7B Coding Agent Trained From Scratch for ~$1,000 Beats Claude 3 Opus and GPT-4o on SWE-bench
OfirPress · x · 2026-10-12
The nanoswe team trained a 1.7B coding agent from scratch with only about $1,000 of training compute, reaching 25.5% pass@1 on SWE-bench Verified — beating Claude 3 Opus (15.8%) and GPT-4o (23%) and setting a new record for their open SWE-bench speedrun. A follow-up thread explains how this was achieved.
More from coding & agent
- Meta's GitSwarm: Git repo as shared memory for agent swarms, solving IMO-ProofBench in one run — armand_ruiz · 2026-10-13
- Runlayer launches secured Muse connectors to govern agent access to company data — gabriel1 · 2026-10-13
- AI agents write 3x more code but PRs merge 4x slower: ZeroSlop CI gates strip 94% of AI bloat — Exciting_Item_2732 · 2026-10-13
- Meta's Alexandr Wang shares one-line curl install for Muse Code — alexandr_wang · 2026-10-13
- When is an AI agent worth it? A 4-condition checklist and ROI formula — emmanuelvivier · 2026-10-13
- Meta explains Muse Wonder's Active Memory and Active Focus for long-running agents — AIatMeta · 2026-10-13