Liquid RIPL Engine Cuts AI Inference Latency

The new engine enables sub-50 ms response times for foundation models, aiming to support real-time production workloads.

Updated on Oct. 22, 2026 in Artificial Intelligence

Isometric editorial illustration of a dense array of matte metallic server cooling fins, representing high-speed data processing infrastructure. — Liquid AI has released its RIPL inference engine, enabling sub-50 ms latency for foundation models in production environments like Shopify's search tools. AI Illustration. Upload story photo >

Live Poll

Is faster AI inference more important than other performance metrics for your daily digital services?

Liquid AI has released its RIPL inference engine, a tool designed to serve foundation models within live request paths. Shopify has already deployed the engine, reporting sub-20 ms inference speeds for production search tasks.

Why it matters

Traditional large language model serving often fails to meet latency requirements for real-time applications like payment authorization and search ranking. Liquid RIPL aims to solve this by optimizing the end-to-end request path to keep response times under 50 ms.

Liquid AI reports that its RIPL engine delivers end-to-end responses in under 50 ms, claiming latency up to 28.4 times lower than the Qwen3.5 model running on vLLM. The models support diverse hardware including NVIDIA H100, H200, B200, and RTX PRO 6000 GPUs.

The players

Liquid AI

A startup developing specialized inference engines and foundation models designed for high-speed, low-latency execution.

Shopify

A multinational e-commerce platform that has deployed Liquid AI models into its production environment for search enhancement.

The details

Liquid RIPL optimizes the full request path by utilizing hybrid layer architectures that combine standard GQA (Grouped Query Attention) blocks, which speed up memory access, with convolution blocks for temporal processing. The engine processes models such as the LFM2.5-1.2B-Instruct, which features 16 total layers, and the larger 2.6B-parameter version, which scales to 30 layers. By integrating these specific layer structures, the engine facilitates real-time inference that replaces the slower processing speeds typical of standard LLM serving architectures.

October 22, 2026: Liquid AI released the Liquid RIPL inference engine.

The Tech Race

Liquid AI is positioning its engine against industry-standard serving solutions like vLLM, which are commonly used for deploying LLMs. The push toward sub-50 ms inference marks a competitive shift aimed at making generative models viable for high-frequency infrastructure tasks.

Developers and companies managing real-time search or payment systems can now utilize RIPL to lower latency for small-to-medium foundation models. The technology requires specific hardware support, including NVIDIA H-series or RTX PRO 6000 GPUs to achieve the performance gains.

The takeaway

Liquid AI is attempting to move LLMs from chat-based assistants into the critical path of real-time software infrastructure. Observers should track if these latency claims hold up when tested on public benchmarks against larger parameter models than the ones currently highlighted.