Qwen3.5 Fine-Tuning Guide – Unsloth Documentation
7 hours ago
- #Fine-tuning
- #Qwen3.5
- #Unsloth
- Unsloth enables fine-tuning of Qwen3.5 models (0.8B to 122B) with improved speed and reduced VRAM usage.
- Supports both vision and text fine-tuning, with specific VRAM requirements provided for different model sizes.
- Free Google Colab notebooks available for fine-tuning smaller models (0.8B, 2B, 4B, 9B).
- Recommendation to mix reasoning-style examples with direct answers to preserve reasoning ability.
- Full fine-tuning (FFT) possible but uses 4x more VRAM; QLoRA (4-bit) training not recommended for Qwen3.5.
- MoE models (35B, 122B) support faster training with less VRAM; bf16 setups preferred over QLoRA.
- Vision fine-tuning supported with options to fine-tune specific parts of the model (vision/language layers).
- Guidance provided for saving/exporting models to various formats (GGUF, vLLM) and deployment options.
- Updates and specific setup instructions for transformers version and Unsloth installation highlighted.