Cerebrium

Cerebrium 做的是无服务器 GPU 推理：把模型或服务打成镜像推上去，按秒计费，不碰 Kubernetes，冷启动压在亚秒级。官方给场景是实时语音 Agent、视频模型与 LLM 推理这类对延迟敏感、流量又起伏不定的负载，可以留一个最小常驻实例把冷启动彻底消掉，也能挂持久卷、设环境变量与多环境部署。适合模型上线初期还没摸清流量曲线的团队。 🔗 文档: https://www.cerebrium.ai/docs

Cerebrium runs serverless GPU inference: push a model or service as an image, pay per second, skip Kubernetes entirely, and keep cold starts under a second. Its stated targets are realtime voice agents, video models and LLM inference — latency-sensitive loads with spiky traffic — with an optional minimum warm instance to remove cold starts altogether, plus persistent volumes, environment variables and multi-environment deploys. It fits teams shipping a model before the traffic curve is known. 🔗 Docs: https://www.cerebrium.ai/docs