리퀴드 AI, '추측적 디코딩'으로 VLM 추론 3배 가속

리퀴드 AI가 소형 초안 모델이 여러 토큰을 예측하고 대형 모델이 이를 한꺼번에 검증하는 '추측적 디코딩' 방식을 비전-언어 모델 (VLM)에 적용한 '디스파크 (DSpark)'를 공개했다. 이는 기존에 텍스트 중심으로 쓰이던 추측적 디코딩 기술을 시각 정보 처리가 결합된 멀티모달 영역으로 확장해 추론 속도를 높였다는 점에서 의미가 있다.

AI타임스·Sep 28

Sarvam AI Releases Saaras V4: A Speech-to-Text Model for All 22 Indian Languages and Global English

MarkTechPost·Sep 26

A new, blazing-fast version of the lightweight visual language model 'LFM2.5-VL-3B' has been released, 'LFM2.5-VL-3B-DSpark,' which applies the DSpark draft model to the LFM2.5-VL-3B.

GIGAZINE·Sep 25

AI Industry's GitHub Seeks Sale, Hugging Face Valuation May Exceed $13 Billion

AIBase·Aug 24

小型モデル「LFM2.5」を2倍以上高速化する投機的デコーディング技術DSpark適用版が登場

AI企業のLiquid AIが小型モデルのLFM2.5シリーズに高速化技術の「DSpark」を適用したことを発表しました。DSpark適用版のモデルは性能を維持したまま2倍以上の高速化に成功しています。

GIGAZINE on MSN·Aug 21

A QAD (Quick Adapter) version of the 'LFM2.5,' a compact AI model that runs on smartphones, has been released, featuring reduced memory usage while minimizing performance degradation.

GIGAZINE·Aug 20

MiniCPM5-2B Benchmark: Beats Bigger Rivals by 2.8 Pts

shattered.io·Sep 27

Supersonic Labs Releases Julia 1: A 144.3M-Parameter Open Decision Model That Runs on a CPU

MarkTechPost·Sep 26

軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場

アメリカに拠点を置くAI開発企業のLiquid AIが視覚言語モデル「LFM2.5-VL-3B-DSpark」を2026年9月24日に公開しました。LFM2.5-VL-3B-DSparkは「LFM2.5-VL-3B」に投機的デコード技術「DSpark」を適用したモデルで、性能を維持したまま数倍の高速化を実現しています。

GIGAZINE on MSN·Sep 24

LFM2.5 DSpark草稿模型发布，推理速度最高飙升3. 18 倍

日前，Liquid AI 与 Hugging Face 正式发布了 LFM2.5系列三款核心模型的 DSpark 草稿模型检查点，包括 LFM2.5-1.2B-Instruct、LFM2.5-2.6B 以及 LFM2.5-8B-A1B。通过引入全新的投机解码路径，该技术在完全不改变模型输出质量的前提下，大幅提升了推理吞吐量。 核心性能提升与应用表现 ...

ITBear科技资讯 on MSN·Aug 21

Up to 3.2x Faster Inference with LFM2.5-DSpark

Hugging Face·Aug 20

Meet UPDF: A Lightweight Adobe Alternative Built for the Agentic Era

MarkTechPost·Aug 20

AI Coding Agents for Enterprise: IP Indemnity, Data Residency and 500-Seat Cost Compared

marktechpost.com·Sep 27

Aikido Security Releases Altar-1: An Open-Weight Security Model Pruned From GLM-5.3 to 328 GB

MarkTechPost·Sep 25

LFM2.5-VL-DSpark reduz latência de modelos de visão-linguagem em até 3,13x

Liquid AI acelera o LFM2.5-VL-3B com um drafter de 280M parâmetros: decode até 3,13x mais rápido, sem perder qualidade. Veja como funciona e onde não ajuda.

iMasters·Sep 24

Liquid AI与Hugging Face联手，LFM2.5 DSpark草稿模型推理性能大幅提升

Liquid AI与Hugging ...

ITBear科技资讯 on MSN·Aug 21

A QAD (Quick Adapter) version of the 'LFM2.5,' a compact AI model that runs on smartphones, has been released, featuring reduced memory usage while minimizing performance ...

AI company Liquid AI offers the LFM2.5 series as a compact AI model that can run on low-spec devices such as laptops and smartphones. A new version of the LFM2.5 series has been released as an open...

GIGAZINE·Aug 20

Liquid AI Releases LFM2.5-VL-3B: A 3B Vision-Language Model That Reads Screens, Grounds Objects, and Calls Tools On-Device

MarkTechPost·Aug 13