Unsloth has released comprehensive documentation for fine-tuning Qwen3.5 LLMs, covering the entire model family from 0.8B to 122B parameters. The guide details how to fine-tune these models with support for both vision and text tasks, making Qwen3.5 accessible to developers with varying hardware capabilities.

Unsloth claims to make Qwen3.5 training 1.5× faster while using 50% less VRAM compared to traditional FA2 setups. The documentation provides specific VRAM requirements for different model sizes, with the 35B MoE variant requiring 74GB VRAM for bf16 LoRA fine-tuning. The team offers free Google Colab notebooks for fine-tuning smaller models and provides guidance on preserving reasoning capabilities through example mixing.

The guide covers essential technical details including Mamba Triton kernel compilation, MoE fine-tuning considerations, and export options to formats like GGUF for deployment. It warns against 4-bit training for Qwen3.5 models due to quantization issues and provides specific instructions for vision fine-tuning, multi-GPU setups, and saving fine-tuned models to various inference platforms.