Accelerating vision-language models with LFM2.5-VL-DSpark

Liquid AI has shipped an experimental DSpark draft model for LFM2.5-VL-3B, its vision-language model. The pitch is familiar if you’ve been following speculative decoding: add a drafter, spend a bit more memory, and get a faster model without changing what it says. Here, the company says the drafter adds about 280M parameters, or 8.9% on top of the 3B target.

The speed numbers are the main event. On-device, Liquid AI reports decode speedups of up to 3.13x on an M5 Max with MLX and up to 2.14x with llama.cpp on an M3 Ultra. On an H100, decode goes up to 2.66x faster. End-to-end gains are smaller, which is exactly what you’d expect when only part of the pipeline gets accelerated: up to 2.62x on device and 2.27x on H100.

The drafter itself follows the same basic recipe as the company’s text DSpark models. It taps hidden states from the target model at fixed layers, then uses those states to propose a block of k candidate tokens. For vision inputs, image patches and text tokens are projected into a shared representation before those layers, so the drafter can work with the same hidden-state shape no matter the modality.

Liquid AI says it trained the model on a mixture of vision-language supervised fine-tuning data, weighted toward the workloads it expects the model to serve. After ablations across 3, 4, and 5 layers, it settled on a simplified attention-only drafter with 4 layers and a block size of 9. Training ran for 10 epochs, and the company recommends a block size of 8 or 9 at inference depending on the hardware.

The model is available with day-one support for llama.cpp, MLX-VLM, and SGLang. Liquid AI also says speculative decoding is exact: the target model checks every proposed token, so the greedy output matches the target model alone. That matters because this is not a “close enough” speed trick; it’s a way to make the same model behave better under load.