LiquidAI releases DSpark draft model to accelerate LiquidAI’s LFM2.5-VL-3B vision-language model

LiquidAI announced an experimental DSpark draft model for its open-weight vision-language model LFM2.5-VL-3B, released on Hugging Face. The drafter adds a speculative decoding path that increases the deployed model’s parameter count by 8.9% (280M parameters) while preserving output quality.

The drafter uses the same architecture as LiquidAI’s text DSpark drafters, conditioning on hidden states at fixed layers to draft token blocks. Because image patches and text tokens share a common representation before those layers, the inference algorithm remains unchanged across modalities.

Training followed the DSpark recipe with a vision-language supervised fine-tuning mixture, optimized over block sizes and layer counts. The final drafter has 4 attention-only layers and a block size of 9, trained for 10 epochs with diminishing returns after additional tokens.

On-device results on Apple silicon show decoding speedups from 2.30x to 3.13x and end-to-end latency gains from 1.56x to 2.62x across six vision-language tasks. On an H100 GPU, decoding improves from 2.04x to 2.66x and end-to-end from 1.64x to 2.27x.

Day-one integration is provided for llama.cpp, MLX-VLM, and SGLang. The draft model is available in Safetensors and GGUF formats on Hugging Face under open weights, enabling download, fine-tuning, and deployment without restrictions.