Liquid AI launches RIPL to cut inference latency below 50ms

Liquid AI released RIPL (Real-time Inference Pipeline), a real-time inference engine aimed at enterprise workloads, on October 22, 2026, according to aitimes.com. The company says the product keeps end-to-end response latency under 50 milliseconds.

RIPL connects Liquid AI's own Liquid Foundation Model (LFM) family directly into the request path. It is a commercial product, not open source, and pricing is available through Liquid AI's sales team.

The engine is built for short-output tasks such as search query rewriting, intent classification and fraud detection rather than long chatbot answers. In Liquid AI's own benchmarks, LFM2.5-1.2B on RIPL had 14.4 to 28.4 times lower p95 latency than Qwen3.5-0.8B on vLLM, while LFM2.5-2.6B on RIPL had 5.6 to 7.7 times lower p95 latency than Qwen3.5-2B on vLLM.

Shopify has adopted the LFMs for commerce tasks and achieved under 20 milliseconds of inference time on live production traffic for search improvements, according to the company.

RIPL supports NVIDIA H100, H200 and B200 GPUs and RTX PRO 6000 GPUs. The LFM2.5 models range from 230 million to 2.6 billion parameters.

Enterprises have mainly relied on fast classical machine learning methods such as gradient boosting trees and small embedding models for real-time tasks because of the response latency of large language models.