Apple rilascia un nuovo modello AI si chiama LensVLM e cerca di risolvere il problema dei token in

Apple, traditionally focused on hardware rather than AI model development, has released a new visual language model called LensVLM with 9 billion parameters. The model is built by fine-tuning Qwen 3.5 and is accompanied by a scientific paper titled 'Selective Context Expansion for Compressed Visual Representation of Text,' which introduces an innovative approach inspired by Deep Seek's OCR research. The key innovation addresses a fundamental challenge in large language models: handling extremely large text inputs that consume excessive tokens. Rather than processing text directly, the technique converts text into visual representations (images containing the text), which are then converted into vision tokens. This conversion achieves a 15x compression ratio because vision tokens consume significantly less space than text tokens. Once compressed and input into the model, the system can selectively expand and reason about specific parts of the image to accurately understand concepts and generate appropriate responses. The speaker expresses surprise at Apple's entry into model research, noting that the company has intelligently positioned itself primarily in the hardware space, where it manufactures powerful devices like Mac Studio with M5 and M6 chips capable of running cutting-edge models like Deep Seek. The speaker speculates that Apple may continue selective research and fine-tuning efforts, potentially releasing more papers and optimized models, though it remains uncertain whether Apple will develop a fully competitive open-source model to challenge Chinese and American competitors or maintain its focus on hardware.