Liquid AI's DSpark draft models hit 3.18x throughput on GPU and cut function-calling latency 57%
Hugging Face Blog (Liquid AI)·high signal
LFM2.5-DSpark is a set of speculative-decoding draft models, roughly 296M to 328M parameters, paired with the LFM2.5 1.2B, 2.6B and 8B-A1B targets. Liquid reports up to 3.18x throughput on GPU and 2.87x on-device, with the 2.6B hitting 2.67x on an H100 and 2.27x on an M4 Max MacBook, and function-calling latency down 57% on average. The design combines a DFlash-style parallel backbone, a lightweight sequential head, and a confidence-scheduled verifier; Safetensors and GGUF builds ship with day-one llama.cpp and SGLang support.