Liquid AI launches RIPL inference engine promising sub-50 ms responses

What was announced

Liquid AI unveiled its Real-time Inference Pipeline (RIPL), an inference engine that runs models from its LFM2.5 family, ranging from 230 million to 2.6 billion parameters, inside a live request path. The company says complete responses come back end-to-end in under 50 milliseconds.

The problem it targets

Most serving stacks are built for large batches and long agentic jobs, where seconds per request are acceptable. Search ranking, fraud checks and payment authorization need answers in milliseconds. Liquid AI argues teams default to classical models for these paths because LLM serving is too slow, and says RIPL is meant to remove that trade-off.

Performance claims

According to Liquid's tests on a single NVIDIA B200:

• 1B class, 32 in → 32 out: 34.6 ms (LFM2.5-1.2B) versus 498 ms (Qwen3.5-0.8B), a 14.4x gap.
• 1B class, 256 in → 128 out: 155.8 ms versus 4,431 ms, a 28.4x gap.
• 2B class on longer requests: the gap narrows to 5.6x, and the sub-50 ms target is missed.

All figures are self-reported. LFMs ran on RIPL while baseline models ran on vLLM.

Deployment

RIPL is available for enterprise deployments on NVIDIA H100, H200, B200 and RTX PRO 6000 GPUs. Liquid AI does not disclose the engine's internals; its comparison of the same LFM2.5-350M model on vLLM and RIPL suggests part of the gain comes from the engine rather than the model architecture.