Liquid AI wants small models to feel fast without changing what they output. On August 20 the company released LFM2.5-DSpark, a set of speculative-decoding draft checkpoints for three members of its LFM2.5 family.
Liquid reports throughput gains of up to 3.18x on a single H100 GPU and up to 2.87x on Apple-silicon MacBooks. The drafts target LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and the mixture-of-experts LFM2.5-8B-A1B, each adding about 300 million parameters of drafter overhead.
The technique works by letting a small drafter model propose tokens while the larger model verifies them in parallel, cutting latency without touching the final results. The checkpoints are aimed at developers running LFM2.5 on laptops, edge devices, and single-GPU servers.
For Liquid, the release extends the practical range of its open-weights family, making modest hardware feel like a much bigger deployment.
