# LFM2.5-VL-DSpark

Liquid AI released the LFM2.5-VL-DSpark vision drafter, which reuses its text DSpark architecture: image patches and text tokens are projected into a shared representation, and a 4-layer attention-only drafter predicts blocks of candidate tokens, adding roughly 280M parameters (8.9% on top of the 3B target) and delivering 2.30x to 3.13x faster decoding with 1.56x to 2.62x end-to-end gains on M5 Max, 1.57x to 2.14x and 1.30x to 1.77x on M3 Ultra, and 1.64x to 2.27x end-to-end on H100, shipped with day-one llama.cpp, MLX-VLM, and SGLang integrations.

Liquid AI released the LFM2.5-VL-DSpark vision drafter, which reuses its text DSpark architecture: image patches and text tokens are projected into a shared representation, and a 4-layer attention-only drafter predicts blocks of candidate tokens, adding roughly 280M parameters (8.9% on top of the 3B target) and delivering 2.30x to 3.13x faster decoding with 1.56x to 2.62x end-to-end gains on M5 Max, 1.57x to 2.14x and 1.30x to 1.77x on M3 Ultra, and 1.64x to 2.27x end-to-end on H100, shipped with day-one llama.cpp, MLX-VLM, and SGLang integrations.