Liquid AI launches RIPL, a real-time inference engine for LFM models, with 14.4 to 28.4x lower p95 latency

Liquid AI has released RIPL, a real-time inference engine for its LFM models. On one B200, LFM2.5-1.2B running on RIPL showed 14.4 to 28.4x lower p95 latency than Qwen3.5-0.8B running on vLLM, according to a report from MarkTechPost.

What was announced

RIPL is an inference engine, not a new model. It serves Liquid AI's LFM family, and MarkTechPost's headline describes it as built for sub-50 ms LFMs.

Reported performance

The figures come from the MarkTechPost report:

• Hardware: one B200
• Model on RIPL: LFM2.5-1.2B
• Comparison: Qwen3.5-0.8B on vLLM
• Metric: p95 latency
• Result: 14.4 to 28.4x lower p95 latency for LFM2.5-1.2B on RIPL

The comparison pairs different models on different serving stacks, so the result reflects the combination of model and engine rather than the engine alone.

Pricing, context window and availability details for RIPL are not covered in the report.

What this means

Liquid AI is releasing serving software for its LFM models alongside the models themselves. The reported result uses p95 latency, a tail-latency measure that matters for real-time applications. Because the benchmark compares LFM2.5-1.2B on RIPL against Qwen3.5-0.8B on vLLM, teams evaluating RIPL should test it on their own workloads and hardware before drawing conclusions about the engine itself.