How they use Nebius AI Cloud

CloudThrill uses Nebius for inference demonstrations and infrastructure deployments. Its vLLM-Omni lab served Z-Image, Wan2.2, Qwen3-TTS and NVIDIA Cosmos 3 on a single H100, swapping models on the same GPU. Models ran on the Nebius VM while requests came from a laptop through an SSH tunnel. CloudThrill published the demonstration scripts and announced a conference session featuring the Nebius H100 demo.

CloudThrill also documented vLLM Production Stack and NVIDIA Dynamo deployments on Nebius AI Cloud using Managed Kubernetes and Terraform. Its Dynamo deployment served Qwen3-8B and TinyLlama on one L40S through a shared endpoint, with TLS, Prometheus, Grafana and persistent model caches. In a routing test, it reported that all 20 requests sharing a prompt prefix reached the cache-warm replica. The multimodal article separately described a planned inference gateway on Nebius using vLLM Production Stack, LiteLLM, LLM Guard and token-based rate limiting.

CloudThrill reported end-to-end deployment times of approximately 20 minutes 41 seconds for vLLM Production Stack and 35–40 minutes for Dynamo. The Dynamo stack used private subnets, VPC-Cilium networking with Hubble observability and NGINX exposed through a Nebius load balancer. Its multimodal request menu dispatched and timed calls from a laptop, while model servers were started separately and swapped manually on the GPU.