Hugging Face AI Updates: September 25, 2026

1. A 280M Drafter Model Speeds Up a 3B VLM by Reading Its Hidden States

Liquid AI published LFM2.5-VL-DSpark, a 280M-parameter drafter that accelerates the 3B LFM2.5-VL through speculative decoding, adding 8.9 percent parameter overhead. The drafter taps the target model’s hidden states at fixed layers to propose a block of k candidate tokens, and because image patches and text tokens project into shared representations, the same path handles both modalities. Reported decoding speedups run 2.30x to 3.13x on an M5 Max with MLX, 1.57x to 2.14x on an M3 Ultra with llama.cpp, and 2.66x to 20.4x on an H100, with end-to-end gains of 1.30x to 2.62x. Open weights ship in Safetensors and GGUF with day-one support for llama.cpp, MLX-VLM, and SGLang. Source