TIGER-Lab/FIM-8B
FIM-8B by TIGER-Lab is an 8 billion parameter coding agent foundation model, based on Qwen3-8B, specifically mid-trained with function-aware Fill-in-the-Middle (FIM) data and post-trained on SWE-Lego agent trajectories. This specialized training significantly enhances its performance on code-related tasks, achieving notable improvements of +3.2 points on SWE-Bench-Verified and +5.4 points on SWE-Bench-Lite compared to a standard SWE-Lego reproduction. It is designed for robust code generation and problem-solving within agentic workflows, particularly excelling in automated software engineering tasks.
Loading preview...
Overview
FIM-8B is an 8 billion parameter model developed by TIGER-Lab, designed as a robust foundation for coding agents. It is built upon the Qwen3-8B base model and distinguishes itself through a unique two-stage training pipeline:
- FIM Mid-training: The model undergoes mid-training using function-aware Fill-in-the-Middle (FIM) data from the TIGER-Lab/FIM-Midtraining-400K dataset.
- Post-training: It is then further fine-tuned (SFT) on
resolved-filtered SWE-Lego agent trajectories.
This specialized training methodology, particularly the FIM mid-training, is the key differentiator, leading to substantial performance gains in automated software engineering benchmarks.
Key Capabilities & Performance
FIM-8B demonstrates superior performance in automated code problem-solving:
- SWE-Bench-Verified: Achieves 35.00 points, a +3.20 improvement over the Qwen3-8B + SWE-Lego baseline.
- SWE-Bench-Lite: Scores 32.70 points, marking a +5.40 improvement.
These results highlight its enhanced ability to resolve software issues and generate correct code within an agentic setup, evaluated using OpenHands CodeActAgent and the official SWE-bench harness.
When to Use This Model
- Automated Software Engineering: Ideal for tasks requiring autonomous code generation, debugging, and problem resolution, especially within agentic frameworks like OpenHands.
- Code Agent Development: A strong foundation model for building and deploying advanced coding agents that need to interact with complex codebases and environments.
- Research in FIM and Agent Training: Useful for researchers exploring the impact of function-aware FIM mid-training and agent trajectory post-training on LLM performance for coding tasks.