Liquid AI announced RIPL, an inference engine for placing LFM models in real-time request paths such as search ranking, recommendations, and fraud decisions. It focuses on completing short responses quickly rather than long agent tasks or large batches, and the company claimed full responses in under 50ms even at peak traffic. The published comparison used LFM on RIPL and Qwen on vLLM, changing both the model and the engine. The difference therefore cannot be attributed to the engine alone, nor can the same latency be assumed for outputs of arbitrary length. Adoption requires checking end-to-end latency with one’s own inputs and load.