br1-pist/Qwen3.5-4B-AgentCoder
Qwen3.5-4B-AgentCoder is a 4.5 billion parameter causal language model developed by Bruno Pistone, fine-tuned from Qwen/Qwen3.5-4B. Optimized for complex reasoning, tool calling, and code generation, it features a 32768-token context length. The model underwent sequential fine-tuning followed by Direct Preference Optimization (DPO) to enhance alignment, coherence, and reasoning accuracy, making it suitable for multi-domain reasoning and code generation in languages like Python and JavaScript.
Loading preview...
Qwen3.5-4B-AgentCoder: Enhanced for Agentic Workflows
Qwen3.5-4B-AgentCoder, developed by Bruno Pistone, is a 4.5 billion parameter causal language model fine-tuned from the Qwen/Qwen3.5-4B base. This model is specifically optimized for agentic capabilities, excelling in complex reasoning, tool calling, and code generation tasks. It leverages a substantial context length of 32768 tokens, making it capable of handling intricate prompts and extended interactions.
Key Enhancements and Training
The model's development involved a two-stage process: initial sequential fine-tuning followed by a crucial Direct Preference Optimization (DPO) phase. The DPO stage utilized approximately 2.5K chosen/rejected response pairs, where rejected samples were synthetically generated to represent poor answers, and chosen samples were derived from real conversations. This post-training alignment significantly improves the model's coherence, reasoning accuracy, and overall helpfulness, while reducing hallucinations and verbosity.
Core Capabilities
- Complex Reasoning: Optimized for multi-domain reasoning, including mathematical, logical, and general Q&A tasks.
- Tool Calling: Designed for effective integration into agentic workflows, enabling robust tool use.
- Code Generation: Proficient in generating code for various languages, including Python and JavaScript.
Recommended Use Cases
- Developing AI agents requiring strong reasoning and tool interaction.
- Automating code generation tasks.
- Applications needing robust logical and mathematical problem-solving.
Limitations
It is not recommended for highly sensitive data, domains requiring expert-level specialization, or tasks where full explainability is mandatory. While strong, code generation may occasionally produce minor syntax issues, and tool use accuracy depends on task complexity.