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

The LFM2.5 family from a research group now includes DSpark draft model checkpoints for three models, enabling up to 3.2x faster inference on GPUs and nearly 2.9x on-device without compromising output quality. DSpark uses speculative decoding with a lightweight draft model and verification to reduce latency, especially improving function-calling speed by 57%. The draft models are small (~300M parameters) and trained on diverse data, with open-source support for llama.cpp and SGLang from day one.