AIBRIEF
A draft model is a small helper LLM that proposes tokens the main model then verifies, and Liquid AI's DSpark speculative decoding configuration of that recipe trims function calling latency by 57% on its 2.6 billion parameter LFM2.5 base model.
Liquid AI has released draft model checkpoints for three sizes in its LFM2.5 family, with 1.2B, 2.6B, and 8B-A1B base models, pairing each with a small ~300M-parameter "draft" helper that proposes tokens the main model then verifies, a technique known as speculative decoding. On the vendor's published benchmarks, the DSpark variant of this recipe trims function-calling latency by 57% on the LFM2.5-2.6B base model and lifts throughput by up to 3.18× on GPUs and 2.87× on-device.
The 57% cut is where users actually feel the gain. AI agents (copilots, voice assistants, phone-side helpers) spend much of their response time waiting for the model to emit structured function calls, and that loop determines whether on-device assistants feel responsive. The 3.18× GPU number is the same recipe benchmarked on server-class hardware.
DSpark combines a DFlash-style parallel backbone conditioned on the target model's context, a lightweight sequential Markov-chain head for inter-token dependency, and a confidence-scheduled verifier that prunes low-confidence suffixes. It is the latest step in a maturing speculative-decoding family that includes EAGLE-3 and DFlash, not a one-off.
The numbers are Liquid AI's own. Training used 15 epochs over an SFT/chat/code/function-calling mix, with the best epoch selected by acceptance rate rather than loss. Day-one upstream support is already merged into llama.cpp and SGLang, so the speedup is live for anyone running the stack, not gated behind a vendor API. Independent reproduction is not visible in the source bundle, and the gains are strongest on Liquid AI's own models.