Lego-X/SWE-Lego-Qwen3-32B

TEXT GENERATIONPricing:Input $0.408 / Cached $0.0816 / Output $1.972Concurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jan 5, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

SWE-Lego-Qwen3-32B is a 32 billion parameter language model developed by SWE-Lego, fine-tuned from Qwen3-32B. It is specifically designed for software engineering (SWE) issue resolution, achieving a 52.6% Pass@1 and 58.8% TTS@16 on SWE-Bench-Verified. This model leverages a unique supervised fine-tuning (SFT) recipe, including a 32k task instance dataset, error masking, and a difficulty-based curriculum, making it highly effective for automated software bug fixing and development tasks.

Loading preview...

SWE-Lego-Qwen3-32B: Advanced Software Issue Resolution

SWE-Lego-Qwen3-32B is a 32 billion parameter model developed by SWE-Lego, specifically fine-tuned for software engineering (SWE) issue resolving. It is built upon the Qwen3-32B architecture and utilizes a novel supervised fine-tuning (SFT) approach named SWE-Lego.

Key Capabilities and Differentiators

  • State-of-the-Art Performance: Achieves 52.6% Pass@1 and 58.8% TTS@16 on the challenging SWE-Bench-Verified benchmark, demonstrating strong capabilities in automated software bug fixing.
  • Specialized SFT Recipe: The training incorporates a unique methodology with three core components:
    • SWE-Lego Dataset: A high-quality collection of 32,000 task instances and 18,000 validated trajectories, combining real and synthetic data.
    • Refined SFT Procedure: Employs error masking and a difficulty-based curriculum to enhance action quality and overall performance.
    • Well-trained Verifier: Improves test-time scaling (TTS) for more robust solutions.
  • Open-Source Ecosystem: The entire project, including the dataset, code, and training scripts, is open-sourced, fostering further research and development in software engineering agents.

Ideal Use Cases

  • Automated Bug Fixing: Excels at resolving software issues and generating correct code patches.
  • Software Development Agents: Suitable for integration into autonomous agents designed to interact with codebases and fix problems.
  • Code Generation and Refinement: Can be applied to tasks requiring precise code modifications and understanding of software contexts.

Finetunes
1 models