Liquid AI Boosts LFM2.5 Inference Speed by Up to 3.2x

Liquid AI has launched DSpark draft model checkpoints for three models in its LFM2.5 lineup: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. These lightweight draft models, each containing roughly 300 million parameters, implement a speculative decoding path. This technique trades a small memory increase for a substantial boost in decoding speed. Because the target model verifies every proposed token, the final output remains identical to standard greedy decoding, preserving benchmark accuracy.

The new draft models achieve up to a 3.18x throughput improvement on a graphics processing unit and up to a 2.87x speedup on-device. Testing on an NVIDIA H100 80 GB GPU using BF16 precision with SGLang demonstrated these high-end gains. On-device testing with an M4 Max MacBook Pro running FP16 GGUF weights via llama.cpp and Metal showed that LFM2.5-2.6B can reach speeds of around 140 tokens per second. Furthermore, this model saw its function-calling latency cut by an average of 57 percent in multi-tool scenarios.

Performance gains vary across the different models. While the LFM2.5-2.6B shows dramatic improvements, the LFM2.5-1.2B-Instruct experiences speedup variances of up to 52 percent depending on the underlying text distribution. Meanwhile, the LFM2.5-8B-A1B model achieved only an 18 percent average improvement on-device. This smaller gain is attributed to the current Mixture of Experts implementation in the Metal backend of llama.cpp, where verifying multiple tokens activates more experts and increases weight traffic.

For practitioners, these draft models offer immediate integration via day-one support in SGLang and llama.cpp. The DSpark architecture utilizes a parallel backbone conditioned on context features, a sequential head structured as a Markov chain, and a confidence-scheduled verifier to prune low-confidence suffixes. To train these attention-only draft models, which feature 5 layers and a block size of 9, Liquid AI ran 15 epochs on a diverse dataset spanning chat, code, and function-calling data, selecting the epochs with the highest acceptance rates.