UnslothAI has announced the availability of vision fine-tuning for the Llama-3.2-Vision-11B model on Google Colab, enabling users to finetune vision language models (VLMs) at double the speed and with 50% less VRAM usage, while maintaining accuracy. This update also supports other models including Pixtral, Qwen2 VL, and various Llava variants. Users can expect a performance improvement of 1.3x to 2x faster for each model. Additionally, the latest version of OmniVision-968M has been enhanced based on user feedback, showcasing improvements in art descriptions and complex image handling. These advancements are part of a broader trend in optimizing large language models (LLMs) for efficiency and cost-effectiveness, as highlighted by the recent updates from LLM Compressor, which aims to reduce inference times and costs with minimal accuracy trade-offs.