Liquid AI introduced RIPL (Real-time Inference Pipeline) on October 22. It reports complete generated answers under 50 milliseconds at continuous peak load.

The big change

  • What changed: Liquid times complete generated answers during sustained traffic.
  • Why it matters: First-token speed omits the rest. Operators need completion p95 for their requests.
  • What to watch: Enterprise GPU deployment requires sales contact. Buyers need comparable workload, quality and price evidence.

Liquid reports 14.4–28.4× lower p95 for LFM2.5-1.2B/RIPL against Qwen3.5-0.8B/vLLM, and 5.6–7.7× for its 2.6B pairing against Qwen3.5-2B. Both model and engine change, so RIPL's separate contribution is unresolved. The post does not establish full application latency. Liquid cites Shopify production use and sub-20 ms inference; its post supplies no independent customer measurement.