Liquid AI Launches RIPL, a Real-Time Inference Engine for Sub-50 ms LFMs

Liquid AI has launched RIPL, short for Real-time Inference Pipeline, an engine for serving its Liquid Foundation Models inside live request paths. The pitch: complete generated responses end to end in under 50 milliseconds, holding up under continuous peak traffic. It serves LFM2.5 models from 230 million to 2.6 billion parameters on enterprise Nvidia GPUs including H100, H200, B200 and RTX PRO 6000.

The target is the plumbing of the internet: search ranking, fraud checks, payment authorization, ad query rewriting. Those systems mostly run on classical models like gradient-boosted trees because LLM serving has been too slow. If that changes, generative models move into decisions that happen millions of times a second.

We think this is the right problem to attack. Most AI serving is built for big batches and long agent jobs where a few seconds is fine. A payment check can't wait a few seconds. Liquid's argument, that teams default to older models only because generative serving is slow, is plausible, and RIPL is designed to remove that trade-off.

The evidence is interesting but needs a caveat up front. Liquid reports 5.6 to 28.4 times lower p95 latency than Qwen3.5 running on vLLM, measured on a single B200. The best case was LFM2.5-1.2B at 155.8 milliseconds versus 4,431 for Qwen3.5-0.8B. The worst was its 2.6B model, at 449.1 versus 2,499. Notice that neither of those headline results is under 50 milliseconds; sub-50 ms holds for short requests and fails on longer ones. And all of it is self-reported, on a closed, enterprise-only engine.

The most convincing piece is the same-model comparison. Liquid ran LFM2.5-350M on both vLLM and RIPL, and RIPL's latency line sits well below. That suggests the engine itself, not just Liquid's hybrid convolution-plus-attention architecture, is doing real work. That's a meaningful signal, and one independent testers should try to reproduce.

The production claim also carries weight: Liquid says Shopify is running its models across latency-sensitive commerce workloads, with sub-20 ms inference in search on production traffic. If that holds, it's the best argument in the announcement.

Now the downside. The workloads Liquid lists include payments risk scoring and credit decisioning. Swapping an interpretable tree model for a generative one in a credit decision isn't only a latency question. It's an explainability and fairness question, and regulators care about why someone was declined. Where this leads if it scales is more consequential choices made by models that are harder to audit. We're for faster, smarter infrastructure. We want the high-stakes uses to come with the receipts.

Liquid's LFM2.5 models are hybrids: the 1.2B instruct model has 16 layers, 10 of them double-gated convolution blocks and 6 grouped-query attention blocks; the 2.6B model has 22 convolution and 8 attention blocks across 30 layers. Liquid mapped six real-time workloads to fixed latency budgets, from 20 ms search intent classification to 150 ms credit decisioning and 100 ms image alt text.