DeepSWE-Preview: Open-source RL-trained coding agent built from Qwen3-32B

AI Impact Summary

DeepSWE-Preview is an open-source RL-based coding agent trained from scratch on Qwen3-32B, achieving 59% SWE-Bench-Verified with test-time scaling and setting a new benchmark for open-weight coding agents. The release includes the dataset, code, training logs, and environment setup, enabling teams to reproduce and extend RL-driven SWE agents without vendor lock-in. It demonstrates a scalable training workflow (R2E-Gym, Kubernetes orchestration, 64 H100 GPUs) for long-horizon programming tasks, highlighting the practicality of in-house, autonomic coding assistants. This enables organizations to prototype tailored SWE agents for their codebases, but expect substantial compute, orchestration, and governance overhead to adopt at scale.