MLX-VLM provides inference and fine-tuning capabilities specifically for Vision Language Models (VLMs) and Omni Models on Mac hardware. It leverages Apple’s MLX framework for optimized Apple Silicon performance. The package supports numerous VLM architectures including LLaVA, Florence-2, IDEFICS, Molmo, PaliGemma, and Pixtral, and includes advanced features like speculative decoding, vision feature caching, KV cache quantization, and distributed inference.