armand0e/Qwen3.5-9B-Coder

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

armand0e/Qwen3.5-9B-Coder is a 9 billion parameter causal language model, finetuned from Qwen/Qwen3.5-9B, with a 32768 token context length. This experimental model is optimized for code generation and tool use, trained on a diverse mix of traces from various models, including Claude and Kimi datasets. It focuses on training for final answers and tool usage while leaving reasoning capabilities untouched.

Loading preview...

armand0e/Qwen3.5-9B-Coder: Code-Optimized Finetune

This model is an experimental 9 billion parameter finetune of Qwen/Qwen3.5-9B, designed with a substantial 32768 token context length. Developed by armand0e, it focuses on enhancing code generation and tool-use capabilities through a unique training methodology.

Key Capabilities & Training

  • Code Generation & Tool Use: The model was specifically trained to excel in generating code and utilizing tools, with training focused on final answers and tool interactions.
  • Diverse Training Data: It leverages a mix of traces from various models and datasets, including "TeichAI/claude-4.5-opus-high-reasoning-250x", "armand0e/kimi-k2.6-claude-code-traces", and "armand0e/minimax-m3-claude-code-traces", among others.
  • Reasoning Preserved: The finetuning process explicitly left the base model's reasoning capabilities untouched, suggesting a focus on practical application rather than foundational reasoning shifts.
  • Efficient Finetuning: The model was trained using Unsloth and Huggingface's TRL library, completing the finetuning process in approximately 4 hours.

Good For

  • Developers requiring a model with strong code generation abilities.
  • Applications involving tool use and structured outputs.
  • Experimentation with models finetuned on diverse, high-quality code and reasoning traces.