Liquid AI Releases LFM2.5-VL-DSpark: An Experimental Draft Model for Faster Vision-Language Inference

Liquid AI has released LFM2.5-VL-DSpark, an experimental draft model designed to accelerate the LFM2.5-VL-3B vision-language model (VLM) via speculative decoding. This addition introduces a minimal memory overhead of 280M parameters (an 8.9% increase over the 3B target) while delivering significant speed improvements without compromising output quality.

Performance Benchmarks

The draft model achieves substantial inference speedups across different hardware configurations, evaluated on six diverse vision-based tasks including general VQA, image captioning, and complex reasoning:

• On-device (Apple Silicon): Using MLX on an M5 Max, decoding speeds increase by 2.30x to 3.13x, with end-to-end latency improvements of 1.56x to 2.62x. On an M3 Ultra using llama.cpp, decoding improves by 1.57x to 2.14x.
• GPU (H100): Decoding runs 2.04x to 2.66x faster, with end-to-end gains of 1.64x to 2.27x.

Architecture and Implementation

The DSpark drafter utilizes a simplified attention-only architecture with 4 layers and a recommended block size of 8 or 9. It operates by capturing the target model's hidden states at fixed tapped layers and conditioning on them to draft candidate tokens. Because image patches and text tokens are projected into a shared representation, the drafter handles multimodal inputs with identical vector dimensions, keeping the inference algorithm consistent with text-only models.

Integration and Limitations

Day-one support is available for llama.cpp, MLX-VLM, and SGLang. While speculative decoding significantly speeds up the generation phase, it does not accelerate the vision encoding or prefill stages. Consequently, on edge devices where prefill constitutes a larger portion of latency due to lower compute power, the end-to-end speedup is capped by Amdahl's law, resulting in more modest overall gains compared to pure decode speedups.