Liquid AI launches RIPL low-latency inference engine
Liquid AI has released RIPL, an inference engine it says returns complete responses from its small foundation models in under 50 ms, aimed at search, fraud and payment workloads that previously relied on non-generative methods.
Liquid AI announced Liquid RIPL (Real-time Inference Pipeline), a serving engine built to run its Liquid Foundation Models (LFMs) directly in the path of a live request. The company says RIPL returns complete generated responses end-to-end in under 50 ms, and keeps doing so under continuous traffic at peak load.
The pitch is aimed at enterprise real-time systems such as search ranking, recommendations, fraud decisions and payment authorization. Liquid says most current serving stacks are designed for large batches and agentic workflows where a few seconds per request is acceptable. Real-time systems have instead relied on faster but less accurate methods such as gradient-boosted trees or small embedding models.
In Liquid's own benchmarks across two workloads, LFM2.5-1.2B on RIPL showed 14.4–28.4× lower p95 latency than Qwen3.5-0.8B served on vLLM. LFM2.5-2.6B showed 5.6–7.7× lower p95 latency than Qwen3.5-2B. The post does not detail the hardware or request shapes behind those figures, and the comparison pairs different models with different serving engines.
Liquid also cites Shopify, which it says is deploying LFMs in production across latency-sensitive commerce workloads. In search enhancement, Liquid says LFMs reach sub-20 ms inference while using about 50% fewer parameters than alternatives and running 2–10× faster without losing quality. The post does not make clear whether Shopify runs RIPL itself.
RIPL is available now for enterprise LFM deployments on NVIDIA GPUs, including the H100, H200, B200 and RTX PRO 6000. Liquid did not disclose pricing.