Publication date: September 25, 2026
Liquid AI released LFM2.5-VL-3B-DSpark on September 24 as an open-weight speculative-decoding model for its 3B vision-language model. The 279.5M-parameter drafter is available on Hugging Face and accelerates the existing model without changing its target output distribution.
The business case is unusually concrete. Teams running local OCR, screen understanding or visual agents can trade an 8.9% parameter increase for vendor-measured end-to-end gains of up to 2.62x on Apple silicon and 2.27x on an H100.
What DSpark changes in a vision pipeline
LFM2.5-VL-3B-DSpark is not a replacement foundation model. It is a small draft model that runs beside LFM2.5-VL-3B. The drafter predicts blocks of candidate tokens, while the original target model verifies them. Under greedy decoding, Liquid says the resulting text is exactly what the target would have generated alone.
That distinction matters for teams with an already validated vision pipeline. The release aims to improve serving speed without requiring a new quality evaluation of a different base model. The drafter has four attention layers, a block size of nine during training and 279.5 million parameters.
The target model remains LFM2.5-VL-3B, which Liquid released in August for document understanding, OCR, screen and UI understanding, grounding, multi-image input and function calling.
The speedup is large on decode, smaller end to end
| Workload and hardware | Decode speedup | End-to-end speedup |
|---|---|---|
| COCO, M5 Max with MLX-VLM | 3.13x | 2.59x |
| MMMU-Pro, M5 Max with MLX-VLM | 2.93x | 2.62x |
| COCO, H100 with SGLang | 2.66x | 2.27x |
| MMMU-Pro, H100 with SGLang | 2.43x | 1.97x |
These are Liquid AI’s measurements, not independent benchmarks. The tests used 16-bit processing, batch size one and temperature zero. Apple tests used block size eight, while the H100 tests used block size nine.
The gap between decode and end-to-end gains is the key buying fact. DSpark does not accelerate image encoding or prompt prefill. On vision workloads, those stages can consume a large share of wall time, especially on edge hardware. Faster token generation therefore does not translate one-for-one into faster completed requests.
Deployment is open, but runtime support is specific
The model card uses Liquid’s LFM 1.0 license and the release describes the model as open-weight, downloadable, fine-tunable and deployable. Safetensors and GGUF artifacts are available on Hugging Face.
Liquid lists day-one support for SGLang, MLX-VLM and llama.cpp. The model card requires SGLang 0.5.19 or newer and MLX-VLM 0.7.2 or newer. The GGUF checkpoint is the route for llama.cpp.
There is no separate API price because the release is a downloadable draft model rather than a hosted inference product. The relevant cost question is whether extra memory and verification compute are outweighed by lower latency or higher throughput on the team’s actual hardware.
Who should test it, and what could block a switch
Edge AI teams, visual-agent developers and document-processing products already using LFM2.5-VL-3B should care most. The first test should replay a production-like mix of OCR, screenshots, charts and multi-turn image conversations, then compare p50 and p95 end-to-end latency, memory use and requests per dollar.
The main reason not to adopt it yet is workload mismatch. Short outputs, image-heavy prompts or long prefill phases leave less decode work to accelerate. Liquid also labels the drafter experimental, and the published performance evidence comes from its own Pipette infrastructure rather than an independent evaluator.
Teams using quantized deployments should be especially cautious. Liquid says all published inference numbers use 16-bit processing and explicitly leaves acceleration of quantized models outside this release.
What this means
For teams already committed to LFM2.5-VL-3B, DSpark is worth testing now because it promises a speed improvement without changing the target model’s output distribution. Start with the slowest interactive vision workflow, not a synthetic token-throughput test.
For teams choosing a vision model from scratch, this release alone is not a reason to switch. First decide whether LFM2.5-VL-3B meets the accuracy requirement, then test whether DSpark improves total request latency enough to justify its extra memory and runtime complexity.
For more information, visit the official announcement of LFM2.5-VL-3B-DSpark on the Liquid AI website.
