Liquid AI Launches RIPL for Fast LFM Inference

The engine claims to maintain this low latency even when traffic continuously arrives at peak load. Liquid AI positions RIPL as a solution for developers to swap gradient-boosted trees for generative models while adhering to strict latency budgets.

RIPL supports LFM2.5 models ranging from 230 million to 2.6 billion parameters and runs on enterprise deployments using NVIDIA H100, H200, B200, and RTX PRO 6000 GPUs. Benchmarks utilized a single B200 GPU.

Liquid AI's tests indicate RIPL offers 5.6 to 28.4 times lower p95 latency compared to Qwen3.5 on vLLM. Specifically, an LFM2.5-1.2B model achieved 155.8 ms p95 latency versus Qwen3.5-0.8B's 4,431 ms, a 28.4x difference.

The problem RIPL addresses is the latency typically associated with LLM serving stacks, which often target large batches and long agentic jobs where seconds per request are acceptable. For applications like search ranking, fraud checks, and payment authorization, millisecond responses are crucial.

Liquid AI argues that teams have historically defaulted to classical models for these time-sensitive paths due to LLM serving slowness. RIPL aims to eliminate this trade-off by optimizing the entire request-to-response path.

The LFM2.5 models themselves are hybrid architectures. For instance, LFM2.5-1.2B-Instruct features 16 layers with 10 double-gated convolution blocks and 6 GQA blocks, while LFM2.5-2.6B has 30 layers, including 22 convolution blocks and 8 GQA blocks.

Benchmarks conducted by Liquid AI under open-loop Poisson arrivals on a single B200 GPU showed RIPL's p95 end-to-end latency for LFMs to be significantly lower than vLLM across various workloads.