Deploy open-source LLMs
Deploy GLM 5.2, DeepSeek, Qwen 3.7, or any open-source LLM as a production API in minutes.
Production LLM serving on serverless GPUs
Run open-source LLMs with fast cold starts, autoscaling, cached weights, and per-second billing.
Cold starts in seconds
Beam boots containers in seconds, even with big model weights loaded, so scaling up doesn't keep your users waiting.
Explore docsElastic scaling
Endpoints scale from zero to thousands of GPUs the moment traffic spikes, and back down when it stops.
Explore docsOpenAI-compatible API
The managed vLLM integration exposes an OpenAI-compatible API so your existing SDKs, clients, and tooling work by swapping one base URL.
Explore docsAny model, any GPU
Serve Llama, Mistral, Qwen, or your own fine-tuned weights on 4090s through B300s. Switching hardware is a one-line code change.
Explore docsWeights cached in PVs
Mount a storage volume and load weights from disk instead of pulling from Hugging Face on every boot. Storage is not billed yet. Planned: $0.0001 per GB-hour of data stored, with no free allowance.
Explore docsPer-second billing
Pay for the seconds your model is actually serving requests.
Explore docsBeam's infrastructure makes productizing our AI straightforward. We can scale up or down when needed, test new GPU options, and keep our users happy.
Dave Carlton, Founder, Hooktheory
production uptime
of inference requests daily
lower latency after a one-line GPU swap
Deploy an LLM in three steps
Define your app in code
Use the managed vLLM integration (or wrap any inference code in an endpoint decorator) and declare GPU, memory, and image in Python or TS.
Deploy with one command
beam deploy ships it to serverless GPUs. You don't need a Dockerfile, Kubernetes, or YAML.
Call it from any OpenAI client
Point your existing SDK at the endpoint URL. Auth, autoscaling, and telemetry are handled for you.
# 1. Deploy the server
$ beam deploy app.py:llama
=> Deployed 🎉
=> https://llama-3-1-8b-abc123.app.beam.cloud
# 2. Call it from any OpenAI client
from openai import OpenAI
client = OpenAI(
base_url="https://llama-3-1-8b-abc123.app.beam.cloud/v1",
api_key="YOUR_BEAM_TOKEN",
)
response = client.chat.completions.create(
model="meta-llama/Meta-Llama-3.1-8B-Instruct",
messages=[{"role": "user", "content": "Hello!"}],
)Frequently asked questions
Which models can I serve?
Any open-source or custom model: GLM 5.2, DeepSeek, Qwen, or your own fine-tuned weights. Use the vLLM integration for an OpenAI-compatible server, or wrap any inference code in an endpoint decorator.
Do I pay while the endpoint is idle?
No. Endpoints scale to zero when idle and billing is per-second, so you only pay while requests are being processed. If you're latency-sensitive, you can keep containers warm and control exactly how long they stay up.
Am I billed for cold starts?
No. Beam only charges for the time it takes to load your application code. It never charges for server spin-up or container image pulls.
Can I keep a server running 24/7?
Yes. By default apps spin down after each request, but you control the idle timeout with keep_warm_seconds, or you can set min_containers to keep capacity always on. Many teams with latency-sensitive workloads run their LLM servers around the clock on Beam.
Is the API OpenAI-compatible?
Yes, when you use Beam's vLLM integration. Your deployed endpoint speaks the OpenAI API, so existing SDKs and client libraries work by changing only the base URL.
Can I stream tokens back to my users?
Yes. The vLLM integration streams through the standard OpenAI API.
How do I monitor my deployment?
The dashboard shows request volume, latency, container counts, and resource utilization per deployment, with streaming logs for every task. Deployments are versioned, so you can roll back to a previous version as needed.
Can I run this in my own cloud?
Yes. Beam's engine is open source (Beta9), and you can self-host it on your own VPC or connect your own hardware and keep using the same SDK and decorators.
Related use cases
Deploy any Hugging Face model as an API
From model ID to live, autoscaling endpoint with one Python file.
Training & Fine-tuningFine-tune models on GPUs
LoRA and QLoRA fine-tuning on serverless GPUs.
InferenceServe embedding models for your RAG stack
Host embedding and reranking models on your own endpoints.
Give your agents their own machines
Sign up, install the CLI, and launch a sandbox, VM, or GPU from your terminal or your coding agent.