YFarmX logoYFarmXIntelligence for an accelerating worldBriefings
YFarmX logoYFarmX
YFarmX

Liquid AI

LFM2.5-VL-DSpark

Liquid AI’s LFM2.5-VL-DSpark helps its 3B vision-language model write answers faster

Released 24 September 20261 min readLarge Language ModelsLast updated:

LFM2.5-VL-DSpark editorial illustration

Key facts

24 Sep 2026
Released
LFM2.5-VL-3B
Main model
About 280M
Companion size
Up to 2.62×
Measured full-task gain

Liquid AI’s LFM2.5-VL-DSpark helps its 3B vision-language model write answers faster. Released on 24 September 2026, the experimental companion is available to download.

LFM2.5-VL-DSpark is a small companion model that helps Liquid AI’s LFM2.5-VL-3B answer questions about pictures faster. Liquid AI released it on 24 September 2026. It drafts several possible next pieces of an answer, which the main model checks together. This process is called speculative decoding.

The main model checks the draft

The companion has about 280 million parameters and adds about 8.9% to the deployed parameter count. Liquid AI reports support in llama.cpp, SGLang and MLX-VLM, with weights available on Hugging Face.

The speed figures measure particular hardware and settings. In Liquid AI’s tests, generating the answer ran up to 3.13 times faster on an M5 Max. The improvement across the complete task was up to 2.62 times, including the work needed to process the picture and prompt. Those are the company’s measurements.

Picture processing still takes time

Speculative decoding speeds up writing the response. Reading the image and processing the initial prompt still consume time, so the benefit depends on the task and answer length. Liquid AI’s reported tests used 16-bit processing; the published results apply to that setup. The release is experimental, and users should measure their own model, hardware and workload together.

All LLMs →