zai-org/SWE-Dev-32B

SWE-Dev-32B: An Open-Source Software Engineering Agent

SWE-Dev-32B is a 32.8 billion parameter model developed by THUDM, designed as an open-source agent for comprehensive software engineering tasks. Based on the Qwen-2.5-Coder-32B-Instruct architecture, this model is part of the SWE-Dev family, which also includes a 7B variant.

Key Capabilities

• Automated Software Development: SWE-Dev-32B is built to handle various software engineering challenges, including issue tracking, code localization, and test case generation.
• High Performance on SWE-bench: The model achieves a 36.6% solve rate on SWE-bench-Verified, approaching the performance of larger models like GPT-4o. This performance is attributed to a comprehensive pipeline for creating developer-oriented datasets from GitHub repositories.
• Scalability Insights: Research indicates that both training data scaling and inference scaling significantly boost performance on SWE-bench. Higher data quality, combined with reinforcement fine-tuning (RFT), further enhances this trend.
• Inference Scaling Benefits: The solve rate for SWE-Dev models increased from 34.0% at 30 inference rounds to 36.6% at 75 rounds, highlighting the effectiveness of inference scaling.

Good For

• Automating Software Engineering Workflows: Ideal for developers and researchers looking to automate aspects of software development.
• Code Generation and Debugging: Its strong performance on SWE-bench suggests proficiency in code-related tasks.
• Research in Agentic AI: Provides a robust foundation for exploring and developing software engineering agents.