Liquid AI Launches RIPL Real-Time Inference Engine for Sub-50 ms LFM Responses

According to MarkTechPost, RIPL is an enterprise-only, closed engine that supports LFM2.5 models from 230 million to 2.6 billion parameters. Liquid AI said it runs on NVIDIA H100, H200, B200 and RTX PRO 6000 GPUs in enterprise deployments; the benchmarks cited by the company used a single B200. Liquid AI also said latency holds while traffic arrives continuously at peak load, though its published comparisons focus on p95 end-to-end latency.

In Liquid AI's own tests, LFMs running on RIPL showed 5.6 to 28.4 times lower p95 latency than Qwen3.5 models running on vLLM. The widest gap was between LFM2.5-1.2B at 155.8 ms and Qwen3.5-0.8B at 4,431 ms for a workload with 256 input tokens and 128 output tokens. The smallest gap was between LFM2.5-2.6B at 449.1 ms and Qwen3.5-2B at 2,499 ms on the same 256-in, 128-out workload. Liquid AI reported 34.6 ms for LFM2.5-1.2B on a 32-in, 32-out workload at 1,000 requests per second, and 76.4 ms for LFM2.5-2.6B on the same short-request profile.

The sub-50 ms figure does not apply to all requests. Liquid AI's results show that longer prompts and outputs push latency above that threshold; the 256-in, 128-out runs were 155.8 ms for the 1.2B model and 449.1 ms for the 2.6B model. The company's claim is therefore tied to shorter live-path requests and to the model size and workload mix.

Liquid AI said most serving stacks are built for large batches and long agentic jobs, where a few seconds per request is acceptable. It argues that search ranking, fraud checks and payment authorization need answers in milliseconds, and that RIPL is meant to remove the trade-off that leads teams to default to gradient-boosted trees or other classical models for those paths.

The company described RIPL as optimizing the full path from request arrival to completed response, not just the model. In tests that put the same LFM2.5-350M model on vLLM and on RIPL, Liquid AI said RIPL's p95 latency line sat below vLLM's in several workloads, isolating part of the gain to the serving engine. The LFM2.5 models are hybrids: LFM2.5-1.2B-Instruct has 16 layers, including 10 double-gated convolution blocks and 6 GQA blocks; LFM2.5-2.6B has 30 layers, including 22 convolution blocks and 8 GQA blocks.

Liquid AI also mapped six real-time workloads against fixed latency budgets, including adtech query rewriting at 50 ms with LFM2.5-350M, search intent classification at 20 ms, search reranking at 30 ms, payments risk scoring at 50 ms with LFM2.5-1.2B, credit decisioning at 150 ms, and image alt text at 100 ms with LFM2.5-VL-450M. Shopify is deploying LFMs across latency-sensitive commerce workloads, according to Liquid AI; in search enhancement, the company said LFMs run with sub-20 ms inference on production traffic, use about 50% fewer parameters than alternatives and deliver 2 to 10 times faster inference.

The benchmark results are self-reported by Liquid AI, and the RIPL engine is enterprise-only and closed. The comparison also places LFM2.5 models on RIPL against Qwen3.5 models on open-source vLLM, so the latency figures combine model architecture and serving-engine differences; Liquid AI's same-model test with LFM2.5-350M is intended to separate the engine's contribution.