Overview

laion/rl_mixed-struct-step37_terminus-structured is an 8 billion parameter language model based on the Qwen3 architecture, developed by laion. It has been fine-tuned using Reinforcement Learning (RL) over 37 steps with a terminus-structured agent, specifically designed for structured tool interactions. This model excels at tasks requiring the use of external tools like bash, view, edit, create, and search.

Key Capabilities & Performance

• Structured Tool Calls: Optimized for interacting with environments via structured commands, making it suitable for automated development workflows.
• Software Engineering Tasks: Demonstrates strong performance on code-related challenges, achieving a 42% pass@3 on the challenging SWEBench-100 benchmark.
• Context Window: Features a substantial 32K context window (24K input + 8K output), allowing it to handle complex and lengthy problem descriptions or codebases.
• RL Training: Leverages the rloo-n training method with the BenSkyRL + Harbor framework, building upon the laion/r2egym-nl2bash-stack-bugsseq-fixthink-again base model.

Use Cases

This model is particularly well-suited for:

• Automated software development and bug fixing.
• Code generation and modification requiring tool interaction.
• Tasks involving command-line execution and file system manipulation.
• Research into RL-driven agents for complex, structured environments.