Liquid AI unveils dspark draft model for vision-language models, boosting decoding speed by up to 3.13x

Liquid AI has released LFM2.5-VL-3B-DSpark, an experimental speculative-decoding draft model for its vision-language model, LFM2.5-VL-3B. The model adds approximately 280 million parameters and accelerates decoding without altering the model’s output. It achieves up to 3.13x faster decoding on Apple silicon and 2.66x on an NVIDIA H100. The model is available on Hugging Face in Safetensors and GGUF formats, with support in SGLang, MLX-VLM, and llama.cpp. Liquid AI labels the release as experimental, under the LFM Open License v1.0, allowing free commercial use for companies with under $10M in annual revenue. DSpark, based on Liquid AI’s text-model DSpark drafters, reads hidden states from multiple layers and predicts the next k tokens. The drafter operates independently of modality, treating text and image patches as tensors. It uses a simplified attention-only model with 4 layers and a block size of 9, recommended for inference. Training data covered common vision-language tasks for 10 epochs, conducted exclusively on AMD hardware. Evaluation on the MMSpec benchmark showed performance across six task types, with results varying by task and hardware.