Definition
Plain language
An open-source software-engineering agent benchmark and reference implementation.
As stated in the literature
An open-source agentic-coding system used as a comparison baseline in software-engineering agent evaluations, including 32B-parameter variants on SWE-bench Verified.
Also called: OpenSWE-32B
Why it matters: Shared open baselines on SWE-bench keep claims grounded by making comparisons reproducible end-to-end.
For example, a 32B OpenSWE variant might be used as the head-to-head reference on SWE-bench Verified when reporting new agent results.
Heard on the show
“There's OpenSWE-32B, which scores 62.”Episode 047 — When Agent Benchmarks Lie: The Harness Problem in Open-Source AI