Liquid AI ships DSpark draft models claiming 3.18x faster LFM2.5 inference
Liquid AI released ~300M-parameter DSpark draft checkpoints for three LFM2.5 models, claiming up to 3.18x throughput on an H100 and 2.27x on an M4 Max MacBook Pro (61 to 139 tok/s), with speculative decoding leaving greedy output identical to the target model. Function-calling latency drops 57% on average for LFM2.5-2.6B, putting on-device agentic loops in reach of a laptop — though these are vendor benchmarks on Liquid's own blog, with day-one llama.cpp and SGLang support still landing via open PRs.
Source: huggingface.co ↗
For LFM2.5-2.6B, speedup on the MacBook is especially noticeable, as it pushes the interactivity level a user can enjoy far beyond the throughput offered by most proprietary cloud models (around ~140 tok/s, depending on the dataset).
Why this matters
- → Speculative decoding cuts inference latency 3.18x on H100, 2.87x on MacBook—narrowing gap between local and cl
- → Function-calling drops 57% latency on LFM2.5-2.6B, making agentic loops viable on-device.
- → Day-one llama.cpp + SGLang support lowers friction for adoption in production stacks.