Liquid RIPL：毫秒级实时智能引擎发布

Today we're announcing Liquid RIPL (Real-time Inference Pipeline), a low-latency inference engine built to put Liquid Foundation Models (LFMs) directly in the path of a live request. It returns complete generated model responses end-to-end in under 50 ms, and it keeps doing so while traffic arrives continuously at peak load.

Most of today's serving infrastructure is built for a different job: general purpose models, large batches, and long running agentic workflows where a few seconds per request is acceptable. Enterprise real-time workloads have far smaller latency budgets. Search rankings, product recommendations, fraud decisions, payment authorizations, and other real-time workloads need an answer in milliseconds. Even small delays add friction to the user experience, like a slower checkout that hurts conversion. That's why these systems have traditionally used fast, but less accurate, methods like gradient-boosted trees or small embedding models.

Liquid’s RIPL changes that trade off. It’s built for real-time constraints, delivering faster, more responsive generative AI experiences that improve how customers search, discover, transact, and interact. 

Results across real-time workloads

In Figure 1, we highlight a few different use cases for small LLMs in real time enterprise workflows, enabled by low-latency LLM inference.

Across the two workloads in Figure 2, LFM2.5-1.2B on Liquid RIPL has 14.4–28.4× lower p95 latency than Qwen3.5-0.8B on vLLM. LFM2.5-2.6B has 5.6–7.7× lower p95 latency than Qwen3.5-2B.

Proven in production

Shopify is deploying LFMs in production across latency-sensitive commerce workloads, where fast and reliable inference is critical to delivering seamless user experiences under high request volumes.

One example is search enhancement, where LFMs enable sub-20ms inference on real production workloads. LFMs using ~50% fewer parameters outperform alternatives while delivering 2–10× faster inference, without compromising quality.

This is exactly where efficient, task-specialized models shine, especially when paired with high-performance inference engines like Liquid's RIPL to deliver extremely low end-to-end latency and high throughput at scale.

Explore the numbers

Pick a GPU and a request shape to estimate end-to-end latency for an LFM2.5 model as the request rate rises.

Get started

RIPL is available today for enterprise deployments of LFMs on NVIDIA GPUs (such as H100, H200, B200, or RTX PRO 6000). If you have a request path with a hard latency budget, we'd like to hear about it.

Talk to our team

Citation