🔍 Read the full analysis: A Faster Path For Vision-Language Models With LFM2.5-VL-DSpark on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Liquid AI released LFM2.5-VL-DSpark, an experimental 280M-parameter drafter for speculative decoding with its LFM2.5-VL-3B model. The company reports decoding speedups of up to 3.13x on Apple silicon and 2.66x on an Nvidia H100, while end-to-end gains are lower; the results have not been independently verified.
Liquid AI has released LFM2.5-VL-DSpark, an experimental draft model designed to speed up its LFM2.5-VL-3B vision-language model, as detailed in the original analysis. The company reports decoding gains of up to 3.13x on an Apple M5 Max and up to 2.66x on an Nvidia H100, with day-one support in llama.cpp, MLX-VLM and SGLang. Those figures are company benchmarks, and independent results are not yet available.
The add-on contains about 279.5 million parameters, which Liquid AI says adds 8.9% to the target model’s parameter count. It is available on Hugging Face in Safetensors and GGUF formats. The company describes the drafter as experimental; it has not announced whether or when it might become a stable release.
DSpark uses speculative decoding: a smaller model proposes blocks of tokens, then the larger target model checks those proposals. Under greedy decoding, Liquid AI says the target’s verification preserves the same output as generation without the drafter. The new model uses four attention-only layers and a block size of nine. Its components include a decoder stack, hidden-state projection and Markov head.
Liquid AI says it tested the model using the MMSpec protocol across six vision tasks, including general and text-based visual question answering, image captioning, chart questions, complex reasoning and multi-turn conversation. On an M5 Max with MLX, the company reports decoding gains of 2.30x to 3.13x and end-to-end latency gains of 1.56x to 2.62x. On an M3 Ultra with llama.cpp, reported gains are 1.57x to 2.14x for decoding and 1.30x to 1.77x end to end. For the H100, it reports end-to-end gains of 1.64x to 2.27x, using a DSpark block size of eight.
Faster Local Vision Model Responses
The release targets a delay users can notice when running vision-language models locally. These systems process image information as well as text, and the visual input adds work before token generation begins. A small drafter could reduce the time spent producing output, making a 3B-class model more responsive on consumer hardware if the reported gains hold in use.
The distinction between decode and total latency matters. DSpark speeds token decoding, but does not remove the work of encoding an image or processing the prompt. Liquid AI’s reported end-to-end gains are smaller than its decode gains, especially on the M3 Ultra results. The integrations also matter for adoption: support in existing tools can let developers try the method without building a separate inference stack.
Products Worth Considering
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Apple Watch Ultra 4 GPS + Cellular, 49 mm
Amazon Kindle Paperwhite Signature Edition (2024), 32 GB, Metallic Black
As an affiliate, we earn on qualifying purchases.
From Text Drafters to Vision
Speculative decoding is a method for accelerating text generation: a smaller drafter proposes several tokens, and the target model verifies them together. Accepted proposals can reduce the number of sequential generation steps. The technique is intended to preserve the target model’s output under greedy decoding because the target model checks the proposals.
Liquid AI previously applied its DSpark approach to text-only LFM2.5 models. The vision release carries the method to a multimodal target. According to the company, image patches and text tokens enter a shared representation before the layers whose hidden states the drafter uses. That lets the drafter use vectors of the same dimensions for either modality, while retaining the inference algorithm used by the text models.
The drafter was trained on a mixture of vision-language supervised fine-tuning data weighted toward expected serving workloads, according to Liquid AI. The company has not published details of that mixture. It says acceptance improved over ten training epochs before gains diminished, and that ablations compared models with three, four and five layers before selecting four.
“It adds a speculative decoding path that trades a minimal increase in memory footprint for a larger speedup without changing output quality.”
— Liquid AI
Benchmark Results Need Outside Checks
The published speed figures are Liquid AI’s own measurements, not independent benchmark results. Performance may vary with hardware, prompt composition and image resolution. The company has not reported how the drafter performs on vision tasks outside its evaluation mix, or given runtime memory measurements beyond the parameter count.
It is also unclear how the drafter affects sampling-based generation, which does not use greedy decoding, and the training data mixture has not been detailed. A GPU benchmark range in the source material is written as “20.4x to 2.66x,” a range that appears internally inconsistent. It may contain a typographical error, but Liquid AI has not clarified the figure; it should not be treated as a reliable speedup value.
The release is labeled experimental, with no public schedule for a stable version. The reported gains establish what Liquid AI measured on its stated setups, not how much faster users will see the model run across other devices and workloads.
Framework Support and Replication
Users can download the model from Hugging Face. The integrations depend on public framework changes: the source material identifies SGLang PR #40651, llama.cpp PR #29339 and MLX-VLM PR #2280. Availability may depend on using builds that include the relevant DSpark support.
The next useful evidence will come from independent testing across different prompts, images and hardware, alongside clearer reporting on memory use and the inconsistent GPU figure. Liquid AI has not said when it will clarify that number or whether it plans to expand the evaluation data. Until then, the headline speedups remain vendor-reported results for the tested configurations.
Key Questions
What is LFM2.5-VL-DSpark?
It is an experimental 280M-parameter draft model that proposes tokens for Liquid AI’s LFM2.5-VL-3B model to verify as part of speculative decoding.
Does it change the model’s answers?
Liquid AI says greedy decoding produces the same output because the target model verifies proposed tokens. The source material does not establish how sampling-based generation is affected.
How much faster is it?
Liquid AI reports decoding gains of up to 3.13x on an M5 Max and 2.66x on an H100. Its reported end-to-end gains are lower, ranging from 1.30x to 2.62x on the listed setups.
Have the results been independently verified?
No independent benchmark results are included in the source material. The reported figures come from Liquid AI, and results may vary by device and workload.
Primary source: Hugging Face · via ThorstenMeyerAI.com
