Liquid AI introduces RIPL, a low-latency inference engine for real-time applications

Liquid AI has unveiled RIPL, a real-time inference engine designed to deliver sub-50 ms latency for its Liquid Foundation Models (LFMs). The system enables end-to-end response generation within milliseconds, maintaining performance even under peak traffic loads. Developers are encouraged to replace gradient-boosted trees with generative models without compromising latency. RIPL optimizes the entire inference pipeline, from request arrival to response completion. Benchmarks show that RIPL outperforms vLLM in several workloads, with the LFM2.5-350M model achieving lower p95 latency. The LFM2.5 models are hybrid architectures, combining convolution blocks and GQA blocks in varying configurations. Liquid AI claims that Shopify is integrating LFMs into latency-sensitive commerce tasks, achieving sub-20 ms inference with 50% fewer parameters than alternatives. The company also highlights RIPL’s performance in six real-time workloads, emphasizing its efficiency and speed. Additional resources, including model cards and a latency calculator, are available for further exploration.