According to CoinMeta, Liquid AI has released a real-time inference engine, RIPL, aimed at meeting the extremely high speed requirements of business scenarios such as search sorting, ad matching, and payment risk control. This engine is capable of generating results within tens of milliseconds, optimizing the entire process from receiving a request to returning a complete result. Official tests have shown that under the conditions of 1,000 requests per second, with 32 token for both input and output, a LFM2.5 model with 1.2B parameters completes 95% of requests within 34.6 milliseconds, while when Qwen3.5-0.8B is used in conjunction with VLLM, it takes approximately 498 milliseconds, representing a speed difference of 14.4 times.