posttrainllm/qwen3-4b-rest-fused
The posttrainllm/qwen3-4b-rest-fused model is a 4 billion parameter Qwen3-4B variant developed by posttrainllm, specifically optimized for file-operation depth in tool-calling scenarios. This research specialist package, trained with a teacher-free ReST iteration method, demonstrates improved performance in file-system interactions and reduced side effects. It is designed for Mac-local tool-calling research or as a component behind an explicit agentic router, rather than general-purpose planning.
Loading preview...
Qwen3-4B ReST Fused: A Specialized Tool-Calling Model
This model, posttrainllm/qwen3-4b-rest-fused, is a 4 billion parameter variant of Qwen/Qwen3-4B-Instruct-2507, developed by posttrainllm. It was produced through a teacher-free ReST (Reinforcement Learning from Self-Training) loop, focusing on enhancing specific tool-calling capabilities.
Key Capabilities and Differentiators
- Optimized for File Operations: The model significantly improves performance in file-operations depth, achieving 12/12 on a hard gate compared to 9/12 for the stock model. It also drastically reduces unexpected depth side effects from 8 to 0.
- Efficiency in Depth Tasks: Demonstrates faster wall time (148.91s vs 360.50s) and improved decode speed (13.40 tok/s vs 12.10 tok/s) for depth-related tasks.
- Research Specialist: This is a research-oriented package designed for specific tool-calling applications, particularly those involving BFCL/OpenAI-style tool calls.
- Teacher-Free Training: Utilizes a unique training method involving self-generated, checker-passing trajectories and a file-operations depth anchor, without external teacher outputs.
Recommended Use Cases
- Mac-local tool-calling research: Ideal for developers experimenting with tool-calling agents on macOS.
- Agentic router component: Suitable for deployment behind an explicit agentic router where specialized file-operation capabilities are required.
Limitations
While excelling in file operations, the model shows a regression in broader tool-calling breadth tasks, making it unsuitable for general-purpose agentic planning without further adaptation. It is not intended as a direct replacement for general-purpose planners like Pace without re-distillation and evaluation on specific intent envelopes.