Liquid AI Releases LFM2.5-VL-3B-DSpark Model

Liquid AI has officially released its new LFM2.5-VL-3B-DSpark model, bringing speculative decoding to vision-language models (VLMs) and achieving up to 3.13x faster decoding speeds. Labeled as an experimental release, the model ships under the LFM Open License v1.0, permitting free commercial use exclusively for companies with under $10 million in annual revenue. Weights are publicly available on Hugging Face in both Safetensors and GGUF formats.

While standard models generate a single token per forward pass, speculative decoding introduces a small drafter model that proposes multiple tokens ahead. The larger target model then verifies the entire block in one pass, retaining agreed tokens. DSpark follows the design recipe from Liquid AI's text-model drafters, reading hidden states from several layers of the target model to predict the next k tokens.

Stock photo for illustration only, not from the actual event

A key architectural advantage is modality-agnostic design. By the time tokens reach hidden layers, both text and image patches are processed simply as tensors, enabling Liquid AI to reuse the exact same inference algorithm for its vision-language model. The drafter itself is a simplified attention-only model utilizing 4 layers and a block size of 9, with Liquid AI recommending a block size of 8 or 9 depending on hardware, and Apple Silicon runs utilizing a block size of 8.

Training utilized supervised fine-tuning data covering common vision-language tasks over 10 epochs, with all ablations and training executed exclusively on AMD hardware. Evaluation followed the MMSpec benchmark across six task categories: General VQA, Text VQA, Image Captioning, Chart VQA, Complex Reasoning, and Multi-turn Conversation, running at batch size 1, temperature 0, and 16-bit weights.