DeepSWE

A 32B software engineering agent achieving 59% on SWE-Bench-Verified

Read the full write-up

Notion blog post with full details

DeepSWE is a 32B software engineering agent that achieves 59% on SWE-Bench-Verified with test-time scaling (42.2% Pass@1). It tops the SWE-Bench leaderboard for open-weight models.

Results

Model	Parameters	SWE-Bench-Verified
DeepSWE	32B	59.0%

Approach

Trained on top of Qwen3-32B to search, view, and navigate codebases, using RL to improve software engineering task completion.

See examples/harbor_swe for the unified-trainer SWE-bench reproducer.

Released: July 2025