Liquid AI makes LFM2.5 decoding 2.27x faster

Liquid AI released roughly 300-million-parameter DSpark draft checkpoints for three LFM2.5 models, reporting mean decoding speedups of 2.27x on an M4 Max and 2.67x on an H100 without changing greedy output. In the release benchmarks, LFM2.5-2.6B rose from 61 to 139 tokens per second on the MacBook and cut multi-tool latency by 57% on average. Local-model operators should file the result as an inference-stack story—not a new model-quality claim—alongside the archive’s earlier DeepSeek serving work.