Liquid AI releases LFM2.5-VL-DSpark for faster vision-language model inference
Liquid AI announced LFM2.5-VL-DSpark, a draft model for its LFM2.5-VL-3B vision-language model. The model uses speculative decoding to speed up inference by up to 3.13x on-device and 2.66x on H100 GPUs, with end-to-end gains of 2.62x and 2.27x, respectively. It adds 280M parameters (8.9% increase) but maintains output quality, according to the company’s benchmarks across tasks like VQA,…
Key points
- LFM2.5-VL-DSpark speeds up vision-language model inference by up to 3.13x on-device and 2.66x on H100 GPUs
- Adds 280M parameters (8.9% increase) to LFM2.5-VL-3B with no output quality trade-off, per Liquid AI
- Supports day-one integration with llama.cpp, MLX-VLM, and SGLang for edge and GPU deployment
The model supports day-one integration with llama.cpp, MLX-VLM, and SGLang, targeting edge and GPU deployments. Liquid AI emphasizes its open-weight approach, allowing unrestricted fine-tuning and deployment. The release aligns with the lab’s goal of AI running across devices, from base models to specialized variants like audio and vision.
Model pages: LFM2.5-VL-DSpark → · LFM2.5-VL-3B-DSpark →
The story so far
5 episodes →- Liquid AI releases LFM2.5-VL-DSpark for faster vision-language model inferencethis story
Accelerating vision-language models with LFM2.5-VL-DSpark
Hugging Face · 24 September 2026
Loading the full article…
This text was published by Hugging Face. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
2sourcesThe headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Introducing Mistral Large 4 · 7 src
- OpenAI boosts GPT-6 speed by 50% in 28 Days update · 1 src
- Reflection AI debuts Beam, a 501B open-weight model rivaling GLM 5.2 with 3-4x less compute · 10 src
- Author tests four AI assistants on flawed forecasting scenarios · 1 src
- Anthropic prompts users to share voice data to enhance AI models · 8 src
Comments
via GitHub Discussions