Liquid AI has released DSpark draft checkpoints for three LFM2.5 models to accelerate decoding without changing the target model's output distribution.
The speculative path uses a lightweight draft model to propose tokens and lets the target verify a block in one forward pass. Liquid reports throughput gains up to 3.18 times on GPU and 2.87 times on-device, with integrations released for llama.cpp and SGLang.
The systems lesson is hardware-aware verification rather than a free model upgrade. Acceptance rate, batch capacity, memory overhead, and target workload determine the realized gain. Teams should reproduce throughput and quality parity on their own prompts and serving configuration before adopting the headline multiplier.