Liquid AI's RIPL claims sub-50 ms responses, 5.6x to 28.4x faster than Qwen3.5

Liquid AI has launched RIPL, an inference engine for its LFM2.5 models, and says it returns full responses in under 50 milliseconds even at peak load. In the company's own benchmarks, p95 latency came in 5.6 to 28.4 times lower than Qwen3.5 running on vLLM. Those are the figures that matter for any product where a user is waiting on a reply.

Latency is a cost as much as a feel. A customer service bot, a fraud check or a checkout recommendation has a budget measured in milliseconds, and a slow tail (the p95, the slowest one in twenty requests) is what users remember. Liquid's pitch is that a smaller model served through a tuned engine can hit that budget without the hardware bill that comes with larger models.

The catch is that the comparison is the company's own, run on its own setup, against one rival configuration. Nothing here is independently audited, and the source does not describe the hardware, prompt lengths or batch sizes used, all of which can swing latency numbers widely. Teams should expect to rerun the test on their own workloads before reading much into the multiples.

Still, for engineering leads building real-time agents, voice interfaces or on-device assistants, RIPL is worth a trial. If the numbers hold up on production traffic, the case for paying for big models on every request gets harder to make.