Liquid AI · 2026-08-20 · notable
LFM2.5-DSpark — Liquid AI's draft models decode up to 3.18x faster
Liquid AI released DSpark draft models for three LFM2.5 checkpoints. Each is about 300M parameters and speeds up decoding through speculative decoding without changing the output: 2.67x on an H100 for LFM2.5-2.6B.
Three ~300M draft models that make Liquid AI's small on-device models decode two to three times faster, with identical output.
Key specs
| Speedup (lfm2.5 2.6 b on h100) | 2.67x |
|---|---|
| Function calling latency on bfcl | -57% |
What is it?
DSpark adds speculative decoding to the LFM2.5 family. Liquid AI shipped draft checkpoints for LFM2.5-1.2B-Instruct (295.7M parameters), LFM2.5-2.6B (327.7M) and LFM2.5-8B-A1B (327.7M). Because the drafter only proposes tokens that the target model then verifies, Liquid AI notes the emitted sequence is identical to baseline greedy decoding by construction, so benchmark accuracy does not move.
How does it work?
Each drafter stacks three pieces: a DFlash-style parallel backbone conditioned on context features that emits several candidate tokens at once, a lightweight sequential head that models the dependency between those tokens as a Markov chain, and a confidence-scheduled verifier that prunes low-confidence candidates before the target model checks them. Mean acceptance out of 10 drafted tokens runs 4.81 for the 2.6B target, 5.02 for the 1.2B and 6.95 for the 8B-A1B.
Why does it matter?
Small on-device models are latency-bound, not quality-bound, and this is a free speed win for anyone already running LFM2.5. On an H100 through SGLang at batch size 1, LFM2.5-2.6B goes from 323 to 864 tokens per second, and LFM2.5-1.2B from 656 to 1,384. On an M4 Max MacBook through llama.cpp with Metal, LFM2.5-1.2B goes from 138 to 350 tokens per second. On the BFCL function-calling set, Liquid AI measures a 57% average latency cut for agent workloads.
Who is it for?
developers running small models on phones, laptops and edge hardware
Try it
https://huggingface.co/LiquidAI/LFM2.5-2.6B-DSpark