Liquid AI’s RIPL targets complete responses in under 50 milliseconds

A model used inside a checkout or search request faces a different time budget from a chat assistant. Liquid AI’s RIPL, announced on 22 October, targets that fast path: the company says it returns complete, short model responses in under 50 milliseconds while handling sustained traffic.[1]

RIPL stands for Real-time Inference Pipeline and serves Liquid Foundation Models. Its intended jobs include search enhancement, recommendations and transaction-related signals. A compact classification or one-line response fits this setting more readily than several paragraphs of generated reasoning.[1]

What the speed ratios compare

The 95th percentile describes the latency below which 95% of requests finish. It makes slower responses visible in a way an average can obscure. But it still depends on the arrival rate and request length: a system that meets its target at one load can build a queue at another.[1]

Because the comparisons change models and software together, they measure the combined configurations. They do not establish that replacing a serving engine alone produces the quoted speedup. Equivalent task quality also matters when comparing a specialised model with another family.[1]

Inference time is one part of a transaction

Liquid cites Shopify’s use of LFMs for production search enhancement, with inference below 20 milliseconds. The claim concerns LFMs; it does not establish that every cited Shopify workload runs on RIPL. Data retrieval, network transit and subsequent application work can add time beyond the inference measurement.[1]

RIPL is available for enterprise deployments on NVIDIA GPUs including H100, H200, B200 and RTX PRO 6000. The deployment choice is therefore concrete: a company can assess a specialised model and serving stack against the latency budget of a real request, rather than treating conversational inference as the only way to use a language model.[1]

Report an error or suggest a correction