Liquid AI Releases LFM2.5-DSpark Draft Models That Deliver Up to 3.18x Faster Decoding Without Changing Model Outputs

Draft models have been released that speed up text generation from larger language models by having a smaller model propose candidate tokens that a larger model then verifies in a single step. The speedup comes from using speculative decoding, where a roughly 300-million-parameter drafter proposes blocks of nine candidate tokens and the target model checks them together rather than generating one token at a time. On high-end GPUs, this approach delivers up to 3.18 times faster decoding while producing identical outputs to running the larger model alone, and the technique provides the largest practical benefit for applications like AI agents that call multiple tools and must wait through reasoning steps repeatedly.