TIGER-Lab/SWE-Next-7B

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Apr 2, 2026License:mitArchitecture:Transformer Open Weights Cold

TIGER-Lab/SWE-Next-7B is a 7.6 billion parameter repository-level software engineering agent fine-tuned from Qwen/Qwen2.5-Coder-7B-Instruct. Developed by TIGER-Lab, it is trained on 3,693 execution-grounded trajectories from real merged pull requests, emphasizing clean repository-level repair and recovery-style debugging. This model excels at real-world software engineering tasks, improving pass@1 on SWE-Bench Verified and SWE-Bench Lite with a 32,768 token context length.

Loading preview...

SWE-Next-7B: A Specialized Software Engineering Agent

SWE-Next-7B, developed by TIGER-Lab, is a 7.6 billion parameter model specifically fine-tuned for repository-level software engineering tasks. It is built upon the Qwen/Qwen2.5-Coder-7B-Instruct base model and leverages the unique SWE-Next SFT Trajectories dataset.

Key Capabilities & Features

  • Repository-Level Problem Solving: Designed to handle complex software engineering tasks that span entire code repositories, not just isolated code snippets.
  • Execution-Grounded Training: Fine-tuned on 3,693 trajectories derived from real-world merged pull requests, ensuring practical and verifiable solutions.
  • Focus on Repair & Debugging: Emphasizes clean repository-level repair traces and recovery-style debugging, making it adept at fixing and improving existing codebases.
  • Scalable Data Collection: The underlying SWE-Next methodology enables efficient collection of large-scale executable data, processing 3,971 seed repositories and 102,582 commit pairs to create 2,308 self-verifying instances.
  • Improved Benchmarking: Demonstrates enhanced pass@1 scores on SWE-Bench Verified and SWE-Bench Lite, indicating strong performance in automated software engineering evaluations.
  • Extended Context Window: Features a 32,768 token context length, crucial for understanding and modifying large codebases.

Good For

  • Automated Software Repair: Ideal for agents tasked with identifying and fixing bugs or implementing features within existing code repositories.
  • Code Debugging: Useful for scenarios requiring recovery-style debugging and tracing issues across multiple files.
  • Research in Software Engineering Agents: Provides a strong foundation and benchmark for further development in AI-driven software development.
  • Large-Scale Codebase Interaction: Its extended context window makes it suitable for tasks involving extensive code analysis and modification.

Finetunes
1 models