Up to 3.2x Faster Inference with LFM2.5-DSpark

DSpark draft model checkpoints are now available for the LFM2.5 family of models, enabling significantly faster inference through speculative decoding. This technique uses a lightweight draft model to propose candidate tokens, which the larger target model then verifies in a single, efficient forward pass. The method combines a parallel backbone, a sequential head, and a confidence-scheduled verifier to achieve speedups up to 3.2x on GPUs and on-device hardware. Importantly, this performance gain is achieved without any change to the final output quality under greedy decoding. The release includes benchmarks on various datasets and integrations with frameworks like llama.cpp and SGLang.