SWE-Star-32B Overview
SWE-Star-32B is a 32.8 billion parameter model developed by LogicStar, built upon the Qwen2.5-Coder family of language models. Its core differentiator is its specialized training on the SWE-Star dataset, which comprises approximately 250,000 agentic coding trajectories. These trajectories were distilled from the Devstral-2-Small model using tasks from SWE-Smith, making SWE-Star-32B highly optimized for automated software engineering.
Key Capabilities & Performance
- Agentic Coding Excellence: The model is specifically designed for agentic coding, enabling it to autonomously solve software engineering tasks.
- Superior SWE-Bench Performance: It significantly outperforms both original SWE-Smith models and other prior work on the SWE-Bench Verified benchmark, which is the standard for agentic Python coding.
- High Pass@16 Rates: Achieves a Pass@16 score of 75.5%, indicating strong potential for reinforcement learning applications.
- Tool Use: Evaluated using an OpenHands-like scaffold with XML-based tool calling, utilizing
think, str_replace_editor, execute_bash, and submit tools.
When to Use This Model
SWE-Star-32B is ideal for applications requiring robust, automated code generation and problem-solving in software development. Its strong performance on agentic coding benchmarks makes it a prime candidate for:
- Automated bug fixing and code refactoring.
- Generating code solutions for defined problems.
- Integration into AI-powered development environments and agents.