Numerata Lupine 让 Qwen3-32B 冷启动从 37.76 秒降至 1.11 秒
做自托管推理的团队该看看这个:Numerata 把 Qwen3-32B 冷启动干到 1.1 秒,靠的是直接迁移 GPU 状态而不是重建,存储走 400 Gb/s RDMA。对比传统 vLLM 的 37 秒冷启动,scale to zero 省钱方案这下真可用了。
Numerata 的 Lupine 引擎在 CoreWeave 基础设施上把 Qwen3-32B 在 vLLM 上的冷启动时间从 37.76 秒压缩到 1.11 秒,提速约 34 倍。其机制是把活跃 GPU 状态直接迁移到新 GPU,避免重建模型实例,实现真正的实时迁移。
存储层通常是这类优化的瓶颈,Lupine 因此把存储层搭建在 CoreWeave Dedicated VAST Storage 上,checkpoint 通过 400 Gb/s RDMA 网络在与 GPU 同一数据中心内传输,绕开了跨机房延迟。
冷启动降到秒级以内后,把推理端点缩容到零(scale to zero)变得实用,闲置时段可以完全释放 GPU,按需拉起,显著降低 32B 级模型的自托管推理成本。
A 32B model was cold-started in just 1.1 seconds on our infrastructure.
Numerata’s Lupine engine moves live GPU state onto a new GPU, so nothing gets rebuilt.
Their Qwen3-32B cold start time on vLLM dropped from 37.76 seconds to just 1.11 seconds.
Storage usually eats that win, but Lupine built its storage tier on CoreWeave Dedicated VAST Storage, so checkpoints move over a 400 Gb/s RDMA fabric in the same data center as the GPUs.
That makes scaling an endpoint to zero practical.
Read how they did it: https://t.co/J0ReIUqFn7
This report is categorized under AITOP's "Products" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Qwen3-32B, vLLM, CoreWeave, inference optimization.
Source attribution: curated from CoreWeave (@CoreWeave).