Type: url
Original: https://www.together.ai/blog/deepswe
Status: partial
Summary
A Together AI research blog post (2 July 2025), written with the Agentica team, introducing DeepSWE-Preview, a coding agent trained from Qwen3-32B with reinforcement learning only on 4,500 R2E-Gym software-engineering tasks using the rLLM framework. It describes the environment and sparse test-based reward, Kubernetes-based rollout scaling, the GRPO++ algorithm and its compact filtering technique, and test-time scaling with hybrid verifiers, reporting 42.2% Pass@1 and 59% with test-time scaling on SWE-Bench-Verified, along with emergent behaviours and experiments that did not work.