Deploy

Run LLM inference on GPUs you already have.

Deploy open-weight models to an OpenAI-compatible endpoint. denv is the control plane. You pay the GPU provider directly.

denv is for teams that want to serve an open-weight model themselves: Llama, Qwen, DeepSeek, or a model you trained. You pick a recipe and a place to run it. You get one OpenAI-compatible endpoint.

It is like EKS on top of EC2, but for inference. The GPU cloud is a provider. denv runs the cluster.

What you are not buying

denv does not sell or resell GPUs, and it does not sell per-token inference. The GPU bill stays between you and Lambda, Nebius, AWS, Hetzner, or whoever owns the machine. denv charges for the control plane: a fee per cluster plus a fee per GPU under management. Rates are by email. See pricing.

If you want a hosted token API and you do not want to think about GPUs, that is a different product. The comparisons say how denv differs from Together, Fireworks, Baseten, Modal, RunPod and SkyPilot.

What a deploy is

  1. Create a cluster. Hosted by denv, or self-hosted on Kubernetes you already run.
  2. Add a site. A provider account, or your own servers with one command.
  3. Create an environment from a recipe: one model on one GPU type, served with vLLM or SGLang.
  4. Call the endpoint with any OpenAI SDK.

The product is early. The quickstart is marked launching soon. Write to anup@denv.ai to see what runs today.

What the endpoint has that a demo server does not

  • Health checks and auto-repair when a GPU throws an XID, an NVLink error, or goes silent.
  • Failover to healthy replicas, including another provider.
  • GPU driver install when a node joins, and driver reinstall as a repair step.
  • NVIDIA MIG and Kubernetes DRA when the model does not need the whole card.
  • A real kubeconfig. kubectl and Helm work. You can leave.

Proof, not a slogan

On one public test, one H100 on this stack placed 3rd of 23 providers serving Qwen, and 2.6–3.9× a typical provider in that test. A separate run on one H100 went from 50.8 to 213 tokens per second. Both were one request at a time. Read the benchmark and the 213 tok/s note before you treat either number as a promise.

A longer map of the decisions, from the first API call to a GPU fleet, is in how to run LLM inference. The same cluster, pointed at a job: voice, coding agents, document AI, computer use, video. Where a latency number does not exist, those pages say so.