Key points
- Liquid AI released RIPL, a real-time inference engine for Liquid Foundation Models.
- Benchmarks show up to 28.4x lower p95 latency compared to Qwen3.5 on vLLM.
- The engine supports enterprise GPUs including NVIDIA H100, H200 and B200.
Liquid AI announced the release of RIPL, a specialized inference engine designed to serve Liquid Foundation Models within live request paths. According to MarkTechPost, the system delivers complete generated responses end-to-end in under 50 milliseconds. This performance holds even when traffic arrives continuously at peak load. The tool targets organizations that require millisecond-level responses for critical operations such as fraud checks and payment authorization.
By the numbers
- LFM2.5-1.2B on RIPL achieved 155.8 ms p95 latency versus 4,431 ms for Qwen3.5-0.8B on vLLM.
- The latency reduction represents a 28.4x improvement for the 1.2B parameter model on a single B200 GPU.
- Larger LFM2.5-2.6B models showed a 5.6x latency gain, with p95 at 449.1 ms versus 2,499 ms.
What the figures mean
Liquid AI argues that teams often default to classical models like gradient-boosted trees for real-time paths because standard LLM serving is too slow. RIPL aims to remove that trade-off by optimizing the full path from request arrival to completed response. Tests plotted LFM2.5-350M on both vLLM and RIPL, showing RIPL’s p95 line sitting well below the vLLM line. This isolates part of the gain to the engine itself, not just the model architecture. The LFM2.5 models are hybrids, combining convolution blocks and GQA blocks.
What should you do?
Managed service providers should evaluate whether current latency budgets for client applications allow for generative model integration. Teams usually need to verify that existing hardware, such as NVIDIA H100 or B200 GPUs, can support the new engine without additional procurement. Check whether your monitoring tools can accurately capture end-to-end latency for these specific inference pipelines. Ensure that security and compliance teams are aware of any shifts from classical to generative models in sensitive paths like payment authorization.
Operational impact
When adopting new inference engines, organizations must assess how changes in latency profiles affect service level agreements. Technical teams should review whether current monitoring frameworks can distinguish between model processing time and network overhead. Procurement departments need to confirm that hardware contracts cover the specific GPU requirements for the new engine. Risk managers should evaluate if replacing classical models with generative ones introduces new compliance considerations for regulated industries.
What to do and how to stay safe: Liquid AI
- Audit current application latency budgets to determine if sub-50 ms targets are feasible for your workloads.
- Verify that your GPU infrastructure matches the supported enterprise deployments listed by the vendor.
- Monitor end-to-end response times during initial testing to ensure consistency under peak load conditions.
- Review security protocols to ensure that switching to generative models does not expose sensitive data in real-time paths.
Step-by-step guide: Shadow AI Explained: 8 Questions Answered for Secure Operations
General security guidance from the BitShield News newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What GPUs does the RIPL engine support?
According to Liquid AI, the engine runs on NVIDIA H100, H200, B200 and RTX PRO 6000 GPUs.
How does RIPL compare to vLLM in terms of latency?
Liquid AI’s tests show RIPL delivers 5.6 to 28.4x lower p95 latency than Qwen3.5 on vLLM.
Which models are compatible with the RIPL engine?
The engine serves LFM2.5 models ranging from 230M to 2.6B parameters.
