TIGER-Lab/FIM-14B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

FIM-14B is a 14.8 billion parameter coding-agent model developed by TIGER-Lab, based on Qwen2.5-Coder-14B-Instruct. It is mid-trained with function-aware Fill-in-the-Middle (FIM) data and post-trained on R2E-Gym agent trajectories, significantly enhancing its performance on coding benchmarks like SWE-Bench. This model excels at agentic code generation and problem-solving, demonstrating improved general capabilities compared to models with only agentic post-training.

Loading preview...

FIM-14B: Enhanced Coding Agent Model

FIM-14B is a 14.8 billion parameter coding-agent model from TIGER-Lab, built upon the Qwen2.5-Coder-14B-Instruct base. Its key differentiator is a novel function-aware Fill-in-the-Middle (FIM) mid-training stage applied before post-training on R2E-Gym agent trajectories. This unique training pipeline significantly boosts its performance in agentic coding tasks.

Key Capabilities & Performance

  • Superior SWE-Bench Performance: FIM-14B achieves +3.0 points on SWE-Bench-Verified and +4.0 points on SWE-Bench-Lite compared to the base model with only R2E-Gym post-training.
  • General Capability Preservation: Unlike typical agentic post-training which can erode general capabilities, FIM mid-training helps recover performance across various benchmarks, showing improvements like +11.1 on LiveCodeBench, +3.9 on ฯ„-bench, and +2.4 on BFCL.
  • Optimized for Coding Agents: The model is specifically designed for use within agent scaffolds like R2E-Gym, providing robust solutions for complex coding problems.

Training Details

The model undergoes a three-stage training process:

  1. Base Model: Qwen/Qwen2.5-Coder-14B-Instruct
  2. FIM Mid-training: Utilizes the TIGER-Lab/FIM-Midtraining-400K dataset.
  3. Post-training: Supervised fine-tuning on R2E-Gym agent trajectories.

Ideal Use Cases

  • Automated Code Generation: Developing intelligent agents for writing and debugging code.
  • Software Engineering Tasks: Solving problems on platforms like SWE-Bench, requiring complex code modifications and problem-solving.
  • Research in Agentic LLMs: Exploring advanced training techniques for coding foundation models.

Finetunes
1 models