독서리서치레이더투자 프레임워크
로그인 / 회원가입
← 읽기로 돌아가기
뉴스간단한 요점1 분 소요DisconfirmAI 편집부

Liquid AI ships a vision draft model, claiming decode speedups above 3x

A draft model adding only 8.9% parameters speeds up vision-model decoding, but end-to-end gains are capped by the vision-encoding stage; figures are vendor-measured.

2026-09-24 22:08 GMT+8

Liquid AI released LFM2.5-VL-DSpark on September 24, a speculative-decoding draft model for its 3B vision-language model LFM2.5-VL-3B. The company reports decode speedups up to 3.13x on device and 2.66x on an H100, with end-to-end latency improving up to 2.62x and 2.27x respectively.

The draft model has roughly 280M parameters, adding 8.9% to the deployed model's size, and ships with day-one support for llama.cpp, MLX-VLM, and SGLang. Speculative decoding is exact: the target model verifies every proposed token, so output is unchanged.

The company itself names the limit: speculation accelerates only decoding, not vision encoding or prefill, and on edge devices those stages take up more of the wall time, so end-to-end gains fall well below decode gains. All speed figures are the vendor's own measurements across six vision tasks.

출처

  • huggingface.co↗
리서치 계속하기 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크