A GPU for inference
Serving a model is a pod with an open port. On Lium you rent a GPU from the CLI, start the vLLM image on it, call the OpenAI-compatible endpoint, and remove the pod when traffic stops. You pay per second while it runs, with no minimum rental.
No L40S is free right now. Reference price $0.38 per GPU-hour. Live L40S prices
Start a vLLM server on an L40S for at most two hours, find its endpoint, remove it
pip install lium.io
export LIUM_API_KEY=sk_...
lium up --gpu L40S --name llm --ttl 2h -y --no-ssh \
--image vllm/vllm-openai:latest --internal-ports 22,8000 \
--cmd="--model Qwen/Qwen3-4B --host 0.0.0.0 --port 8000"
lium ps llm --format json | jq '.[0] | {ip, ports}' # internal 8000 -> public port
curl http://<ip>:<public-port>/v1/models # OpenAI-compatible once the model has loaded
lium rm llm -yWhy rent inference GPUs on Lium
Inference load comes and goes, and per-second billing follows it: a pod costs nothing once it is removed, and a rent with --ttl removes itself when the time is up.
Prices are public and live before you sign up. lium.io/pricing.json has one row per GPU model, and lium.io/pricing shows each model with what is free right now.
Everything works from a terminal, so a script or an AI agent can start a server, check it and remove it with no browser.
Which GPU to pick
Pick the card by the memory your model and its cache need: the L40S has 48 GB, the RTX 5090 32 GB and the RTX 4090 24 GB. lium.io/gpu/rtx-4090 has live RTX 4090 prices.
Prefer Python? lium.io/rent-rtx-5090-python rents an RTX 5090 with the SDK in a few lines.
Reaching the server
Only the ports listed by lium ps --format json are reachable from outside, as a map from the internal port to a public one. Bind the server to 0.0.0.0 on an internal port you exposed, or use lium port-forward to reach it on localhost.
A large model takes minutes to download and load. lium logs llm --follow shows the server output while it starts.
Frequently asked questions
Where is the full vLLM example?
docs.lium.io/developers/sdk/examples/vllm builds a reusable vLLM template with the Python SDK, starts a pod from it and prints the OpenAI-compatible endpoint.
How is a pod billed?
Every pod has an hourly price and is billed per second, with no minimum rental period. Billing stops when the pod is removed. A pod rented with --ttl removes itself when the time is up.
Can an AI agent do this without a browser?
Yes. Signup, API keys, renting, running commands and removal all work from the CLI or the REST API with no browser and no e-mail confirmation. docs.lium.io/developers/agents is the one-page guide.
For AI agents — CLI install, agent skill, headless signup: llms-full.txt