Liquid AI Releases LFM2.5-VL-3B-DSpark: Speculative Decoding for Vision-Language Models With Up to 3.13x Faster Decoding
Liquid AI's draft model cuts vision-language inference latency by 3.13x. Speculative decoding is now shipping in production frameworks.
Why it matters
Speculative decoding—a technique that speeds inference by predicting multiple tokens in parallel—is moving from research to deployed tooling. Liquid AI's 279.5M-parameter draft model demonstrates measurable speedups (3.13x on M5 Max, 2.66x on H100) with identical greedy-decoding output, now supported in llama.cpp, MLX-VLM, and SGLang. This is a capability enabler for practitioners running vision-language models at scale or latency-sensitive edge.
The key facts
10 to knowLFM2.5-VL-3B-DSpark: 279.5M-parameter draft model for speculative decoding
Up to 3.13x faster decoding on Apple M5 Max
2.66x faster on H100 GPU
Identical output under greedy decoding (no quality loss)
Support in llama.cpp, MLX-VLM, and SGLang
Technique: speculative decoding (parallel token prediction for inference acceleration)
LFM2.5-VL-3B-DSpark: 279.5M-parameter draft model
2.66x speedup on H100 GPU
Support in llama.cpp, MLX-VLM, SGLang
Speculative decoding technique (inference optimization, not training)
The story so far
Earlier coverage of this storyline
- Accelerating vision-language models with LFM2.5-VL-DSparkHugging Face Blog
- This story
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Liquid AI has released LFM2.5-VL-3B-DSpark, a 279.5M-parameter draft model that brings speculative decoding to its LFM2.5-VL-3B vision-language model. It delivers up to 3.13x faster decoding on Apple M5 Max and 2.66x on H100, with identical output under greedy decoding. Support ships in llama.cpp,…