Liquid AI ships draft models that cut function-calling latency by 57%

Liquid AI released DSpark draft checkpoints for three models in its LFM2.5 family: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. They add a speculative decoding path that trades a small memory increase for a large decoding speedup, and the company reports output quality is unchanged.

The numbers

Liquid AI reports up to 3.18x throughput improvement on a GPU and up to 2.87x on device. The figure that matters most for agents is a different one: function-calling latency drops by 57% on average for LFM2.5-2.6B. Agent loops spend their time waiting on short tool calls, and halving that wait changes what feels usable on a laptop or a phone.

Support for llama.cpp and SGLang landed upstream at release rather than weeks later, which is the difference between a benchmark post and something you can run this afternoon.

• Up to 3.18x throughput on GPU, up to 2.87x on device.
• Function-calling latency down 57% on average for LFM2.5-2.6B.
• Draft checkpoints for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B.
• llama.cpp and SGLang integration open-sourced upstream on day one.

Why speculative decoding wins here

Decoding is memory-bound, not compute-bound. Most of the latency is streaming weights from DRAM into SRAM rather than doing arithmetic. A small draft model proposes several tokens, the target model verifies them in a single forward pass, and the cost of loading those weights is shared across every token verified instead of paid per token.

DSpark combines three parts: a DFlash-style parallel backbone conditioned on the target model's context features, a lightweight sequential head modeled as a Markov chain between neighboring tokens that raises acceptance at later positions, and a confidence-scheduled verifier that drops low-confidence suffixes when checking them would cost more than it saves.

One training detail worth stealing

The draft models are deliberately plain: attention-only, five layers, a block of nine. Liquid AI ran 15 epochs over the full mix and then picked the epoch with the highest acceptance rate rather than the lowest loss.

That choice is the whole trick. Loss measures how well the draft imitates the target in general; acceptance rate measures how often its guesses actually survive verification, which is the only thing that turns into speed. Optimizing the metric you are paid on rather than the one that is easy to plot is a habit worth carrying into your own work.