LiquidAI has released draft model checkpoints for its LFM2.5 family featuring DSpark speculative decoding, a technique that significantly accelerates inference without compromising output quality. The three models—LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B—achieve up to 3.18x throughput improvement on GPUs and up to 2.87x speedup on-device, while reducing function-calling latency by an average of 57% for the 2.6B model.
DSpark combines three core components: a parallel backbone that produces hidden states for draft tokens in a single forward pass, a lightweight sequential head that adds inter-token dependencies to improve acceptance rates, and a confidence-scheduled verifier that prunes low-confidence predictions. The draft models, each approximately 300M parameters, maintain output parity with baseline models under greedy decoding, ensuring benchmark accuracy remains unchanged despite the dramatic inference speedups.
The models ship with immediate support for llama.cpp and SGLang, enabling deployment on both enterprise GPUs (tested on H100 80GB) and edge devices (M4 Max MacBook Pro). LiquidAI has open-sourced the DSpark integration upstream, making the optimization technique accessible to the broader AI development community.
Key Points
Up to 3.2x faster throughput on GPU and 2.87x speedup on-device through speculative decoding
57% average reduction in function-calling latency for agentic inference applications
No output quality degradation—emitted sequences identical to baseline greedy decoding by construction
Day-one support for llama.cpp and SGLang with open-source DSpark implementation
Successfully deployed on both enterprise accelerators and edge devices with minimal memory overhead