Editorial illustration for Liquid AI's LFM2.5-VL Model Achieves 3.13x Faster Decoding
Liquid AI's New Model Speeds Up Decoding 3x Faster
Liquid AI's LFM2.5-VL Model Achieves 3.13x Faster Decoding
Liquid AI put out a new model this week called LFM2.5-VL-3B-DSpark, a draft model built to speed up its LFM2.5-VL-3B vision-language model without changing what that model actually outputs. The addition is small, about 280 million extra parameters, but the company says it cuts decoding time by up to 3.13x on Apple silicon and up to 2.66x on an Nvidia H100. Weights are already on Hugging Face in Safetensors and GGUF formats, and the release has day-one support in SGLang, MLX-VLM, and llama.cpp, so anyone can pull it down and test it now. Liquid AI is still calling it experimental, and it ships under the LFM Open License v1.0, which caps free commercial use at companies earning under $10 million a year.
The technique behind it is speculative decoding, a method that lets a small "drafter" model guess several tokens ahead while the larger target model checks the guesses in a single pass instead of generating one token at a time. Liquid AI has already used this approach on its text models under the DSpark name. The question with LFM2.5-VL-3B-DSpark is what changes, if anything, when the input includes images instead of just text.
Liquid AI has announced LFM2.5-VL-3B-DSpark, an experimental speculative-decoding draft model for its LFM2.5-VL-3B vision-language model. The drafter adds about 280M parameters and speeds up decoding without changing the model’s output. Liquid AI team reports up to 3.13x faster decoding on Apple silicon and up to 2.66x on an NVIDIA H100.
Why this matters
Speculative decoding usually gets tested on text-only models, so seeing it applied to a 3B vision-language model with a 280M-parameter drafter is worth noting for anyone building on-device multimodal apps. The 3.13x figure on Apple silicon is the headline, but the more useful detail is day-one support in SGLang, MLX-VLM, and llama.cpp, plus weights already up in Safetensors and GGUF. That's the difference between a research curiosity and something a developer can actually swap into a pipeline this week.
We'd still want to see the MMSpec numbers broken out by task type before taking "up to 3.13x" at face value. Speedup claims tied to batch size 1 and temperature 0 tell you about a narrow, low-latency use case, not necessarily what happens under real serving loads with concurrent requests. Liquid AI calling this "experimental" is the right caveat, and it's one worth remembering before anyone rewrites a production stack around it. Still, for teams running VLMs on edge hardware or GPUs where decoding latency is the bottleneck, this is a concrete option to test now rather than a promise to wait on.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Federal Court Upholds Pentagon's Security Risk Label for Anthropic
From Archives Anthropic to Pay Akamai USD 11.6 Billion for Cloud Services
Common Questions Answered
What is the LFM2.5-VL-3B-DSpark model and how does it improve performance?
LFM2.5-VL-3B-DSpark is a speculative-decoding draft model created by Liquid AI to accelerate its LFM2.5-VL-3B vision-language model. By adding approximately 280 million parameters, it achieves up to 3.13x faster decoding on Apple silicon and up to 2.66x faster decoding on Nvidia H100 GPUs, without altering the original model's outputs.
What are the available weight formats and framework support for LFM2.5-VL-3B-DSpark?
The model weights are available on Hugging Face in both Safetensors and GGUF formats for immediate use. The release includes day-one support across multiple frameworks including SGLang, MLX-VLM, and llama.cpp, making it accessible for developers building on-device multimodal applications.
Why is speculative decoding on a vision-language model significant compared to text-only models?
Speculative decoding has traditionally been tested primarily on text-only models, so applying this optimization technique to a 3B vision-language model with a 280M-parameter drafter represents a notable advancement. This demonstrates that the performance benefits of speculative decoding can extend to multimodal AI applications, which is particularly valuable for developers building on-device multimodal apps.
What makes LFM2.5-VL-3B-DSpark practical for real-world development versus just a research project?
Beyond the impressive 3.13x speed improvement on Apple silicon, the practical value lies in immediate framework integration and readily available model weights. The day-one support in SGLang, MLX-VLM, and llama.cpp, combined with weights in Safetensors and GGUF formats, transforms this from a research curiosity into a tool developers can directly integrate into their projects.
Further Reading
- LFM2.5-VL-DSpark: Accelerating vision-language models without changing output quality - Liquid AI
- Liquid AI released LFM2.5-VL-3B-DSpark - DataNorth AI
- LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge - Liquid AI
- LFM2.5-VL-3B for Better and Faster Vision Capabilities for On-Device AI - Hugging Face Blog
- LiquidAI/LFM2.5-VL-3B-DSpark - Hugging Face - Hugging Face
