--- title: Accelerating vision-language models with LFM2.5-VL-DSpark description: Liquid AI released DSpark, a draft model that accelerates their LFM2.5-VL-3B vision-language model via speculative decoding, without changing output quality.... canonical: https://daily.dev/posts/accelerating-vision-language-models-with-lfm2-5-vl-dspark-fkfjk19h6 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Accelerating vision-language models with LFM2.5-VL-DSpark | daily.dev og:description: Liquid AI released DSpark, a draft model that accelerates their LFM2.5-VL-3B vision-language model via speculative decoding, without changing output quality.... og:url: https://daily.dev/posts/accelerating-vision-language-models-with-lfm2-5-vl-dspark-fkfjk19h6 og:image: https://api.daily.dev/og/posts/FKFJk19h6.png og:image:alt: Accelerating vision-language models with LFM2.5-VL-DSpark og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Accelerating vision-language models with LFM2.5-VL-DSpark **[Hugging Face](https://daily.dev/sources/huggingface)** · 5 min read · 0 upvotes · 0 comments ## Summary Liquid AI released DSpark, a draft model that accelerates their LFM2.5-VL-3B vision-language model via speculative decoding, without changing output quality. The 280M-parameter drafter adds only 8.9% to the target model's size while delivering decode speedups up to 3.13x on device (M5 Max with MLX) and 2.66x on H100 GPUs, with end-to-end gains up to 2.62x and 2.27x respectively. It ships with day-one integrations for llama.cpp, MLX-VLM, and SGLang. The architecture shares hidden-state representations across image and text tokens, so the same speculative decoding algorithm used for text models works unchanged for vision. The post also explains why speculation helps decode but not vision encoding or prefill, capping end-to-end gains per Amdahl's law, and provides commands to run the models with each supported framework. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What speedup does the DSpark draft model give for LFM2.5-VL-3B inference? DSpark delivers decode speedups up to 3.13x on device (MLX on Apple M5 Max) and 2.66x on an H100 GPU, with end-to-end latency gains up to 2.62x and 2.27x respectively. The drafter adds only 280M parameters, an 8.9% increase over the 3B target model, and does not change output quality since speculative decoding is exact. _Track speculative decoding gains like this for vision-language models on daily.dev._ ### Which inference frameworks support the LFM2.5-VL-3B-DSpark draft model on launch? Day-one support is available for llama.cpp, MLX-VLM, and SGLang, each requiring a specific build (llama.cpp PR#29339, MLX-VLM PR#2280, SGLang PR #40651). SGLang is launched with a speculative-algorithm flag set to DSPARK and a draft model path, llama.cpp uses a spec-type draft-dspark flag, and MLX-VLM takes a draft-model argument. _Developers picking a serving stack for VLM speculative decoding can follow updates like this on daily.dev._ ### Why does speculative decoding give a smaller end-to-end speedup for vision-language models than for text-only LLMs? Speculative decoding only accelerates the decode phase, not vision encoding or prefill, and VLMs spend more time in prefill because images pass through a vision encoder and produce hundreds of visual tokens before the language backbone processes them. On edge devices with less compute, prefill dominates latency even more, so per Amdahl's law the achievable end-to-end gain is capped by the unaccelerated portion of the workload. _Engineers optimizing VLM inference pipelines can follow these architecture tradeoffs on daily.dev._ ## Similar posts on daily.dev - [Up to 3.2x Faster Inference with LFM2.5-DSpark](https://daily.dev/posts/up-to-3-2x-faster-inference-with-lfm2-5-dspark-nwilr6rly) · Hugging Face · 0 upvotes · 0 comments - [Co-Designing AI Models Using Speculative Decoding for Faster LLM Inference](https://daily.dev/posts/co-designing-ai-models-using-speculative-decoding-for-faster-llm-inference-0zr9mkfau) · NVIDIA Developer · 0 upvotes · 0 comments - [Researchers Found a Way to Make LLMs 8.5x Faster\!](https://daily.dev/posts/researchers-found-a-way-to-make-llms-8-5x-faster--gy4vk6zg4) · Daily Dose of Data Science \| Avi Chawla \| Substack · 0 upvotes · 0 comments - [How to Make LLMs 3X Faster](https://daily.dev/posts/how-to-make-llms-3x-faster-hxypsj0hf) · ByteByteGo · 2 upvotes · 0 comments --- Tags: [#ai-inference](https://daily.dev/tags/ai-inference), [#llama-cpp](https://daily.dev/tags/llama-cpp) [View this post on daily.dev](https://daily.dev/posts/accelerating-vision-language-models-with-lfm2-5-vl-dspark-fkfjk19h6) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Accelerating vision-language models with LFM2.5-VL-DSpark","url":"https://daily.dev/posts/accelerating-vision-language-models-with-lfm2-5-vl-dspark-fkfjk19h6","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/accelerating-vision-language-models-with-lfm2-5-vl-dspark-fkfjk19h6"},"datePublished":"2026-09-24T14:10:18.295Z","dateModified":"2026-09-24T14:10:43.527Z","description":"Liquid AI released DSpark, a draft model that accelerates their LFM2.5-VL-3B vision-language model via speculative decoding, without changing output quality....","image":"https://media.daily.dev/image/fetch/s--WXeu4Ji7--/f_auto,pg_1,q_auto/https://cdn-uploads.huggingface.co/production/uploads/644249b08443bce4c9890a0f/WgfT8N8Xyb6lBxSif7XLO.gif?_a=BAMAMiZY0","thumbnailUrl":"https://media.daily.dev/image/fetch/s--WXeu4Ji7--/f_auto,pg_1,q_auto/https://cdn-uploads.huggingface.co/production/uploads/644249b08443bce4c9890a0f/WgfT8N8Xyb6lBxSif7XLO.gif?_a=BAMAMiZY0","isAccessibleForFree":true,"articleSection":"Hugging Face","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hugging Face","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/f1f55c67d81a4330acf5b90b26b0c8e1","url":"https://daily.dev/sources/huggingface"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/accelerating-vision-language-models-with-lfm2-5-vl-dspark-fkfjk19h6","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-inference,llama-cpp","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hugging Face","item":"https://daily.dev/sources/huggingface"},{"@type":"ListItem","position":3,"name":"Accelerating vision-language models with LFM2.5-VL-DSpark"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/accelerating-vision-language-models-with-lfm2-5-vl-dspark-fkfjk19h6#faq","mainEntity":[{"@type":"Question","name":"What speedup does the DSpark draft model give for LFM2.5-VL-3B inference?","acceptedAnswer":{"@type":"Answer","text":"DSpark delivers decode speedups up to 3.13x on device (MLX on Apple M5 Max) and 2.66x on an H100 GPU, with end-to-end latency gains up to 2.62x and 2.27x respectively. The drafter adds only 280M parameters, an 8.9% increase over the 3B target model, and does not change output quality since speculative decoding is exact. Track speculative decoding gains like this for vision-language models on daily.dev."}},{"@type":"Question","name":"Which inference frameworks support the LFM2.5-VL-3B-DSpark draft model on launch?","acceptedAnswer":{"@type":"Answer","text":"Day-one support is available for llama.cpp, MLX-VLM, and SGLang, each requiring a specific build (llama.cpp PR#29339, MLX-VLM PR#2280, SGLang PR #40651). SGLang is launched with a speculative-algorithm flag set to DSPARK and a draft model path, llama.cpp uses a spec-type draft-dspark flag, and MLX-VLM takes a draft-model argument. Developers picking a serving stack for VLM speculative decoding can follow updates like this on daily.dev."}},{"@type":"Question","name":"Why does speculative decoding give a smaller end-to-end speedup for vision-language models than for text-only LLMs?","acceptedAnswer":{"@type":"Answer","text":"Speculative decoding only accelerates the decode phase, not vision encoding or prefill, and VLMs spend more time in prefill because images pass through a vision encoder and produce hundreds of visual tokens before the language backbone processes them. On edge devices with less compute, prefill dominates latency even more, so per Amdahl's law the achievable end-to-end gain is capped by the unaccelerated portion of the workload. Engineers optimizing VLM inference pipelines can follow these architecture tradeoffs on daily.dev."}}]} ```