Liquid RIPL

Liquid RIPL (Real-time Inference Pipeline) is a low-latency inference engine announced by Liquid AI on October 22, 2026. It is built to put Liquid Foundation Models (LFMs) directly in the path of a live request.

Liquid says RIPL returns complete generated responses end-to-end in under 50 ms, and keeps doing so while traffic arrives continuously at peak load. It targets latency-critical enterprise workloads such as search ranking, product recommendations, fraud decisions and payment authorization, where teams have traditionally used gradient-boosted trees or small embedding models because general LLM serving is too slow.

The radar found this page as a new URL in the liquid.ai sitemap.

Why RIPL

Most serving infrastructure is built for general-purpose models, large batches and long-running agentic workflows where a few seconds per request is fine. Real-time enterprise workloads have budgets of milliseconds, and even small delays hurt things like checkout conversion.

Reported numbers

Comparison	Lower p95 latency
LFM2.5-1.2B on RIPL vs Qwen3.5-0.8B on vLLM	14.4–28.4×
LFM2.5-2.6B on RIPL vs Qwen3.5-2B	5.6–7.7×

In production

• Shopify is deploying LFMs across latency-sensitive commerce workloads, for example search enhancement, with sub-20 ms inference on real production traffic
• Liquid says LFMs with ~50% fewer parameters outperform alternatives with 2–10× faster inference

Hardware

NVIDIA GPUs such as H100, H200, B200 and RTX PRO 6000.

How to use it

• Enterprise deployment: RIPL is available today for enterprise deployments of LFMs; contact Liquid AI ("Talk to our team" on the site)
• Estimate latency: an interactive calculator (liquid.ai/low-latency-calculator-ripl) lets you pick a GPU and request shape and see end-to-end latency for LFM2.5 models as request rate rises
• Announcement: liquid.ai/blog/low-latency

Who it's for

• Search, recommendation, risk and payments teams that need generative models inside a millisecond budget
• Enterprises that want to replace GBDTs or embedding models with small LLMs but are latency-bound