Liquid AI has introduced RIPL, a real-time inference engine designed to achieve sub-50 ms latency for Liquid Foundation Models (LFMs). The engine optimizes the entire request-to-response path, aiming to replace classical models in latency-sensitive applications like search ranking, fraud detection, and payment authorization.
Internal benchmarks comparing the LFM2.5-350M model on vLLM versus RIPL show significantly lower p95 latency for the latter. Tests were conducted on a single B200 GPU using open-loop Poisson arrivals. RIPL outperformed baseline models, including MiniCPM5 and Granite 4.2, which were tested on the vLLM framework.
The LFM2.5 model series features hybrid architectures. The 1.2B-Instruct version contains 16 layers, including 10 double-gated convolution blocks and 6 GQA blocks. The 2.6B version consists of 30 layers, comprising 22 convolution blocks and 8 GQA blocks.
Shopify has begun deploying these models for commerce workloads. In search enhancement tasks, the company reports inference times under 20 ms on production traffic. According to Liquid AI, these models utilize approximately 50% fewer parameters than alternative solutions while delivering 2 to 10 times faster inference speeds.
