Skip to main content
Run Beam yourself: serve its open weights on your own compute, or use Beam through a cloud provider, hardware partner, or inference provider.

SGLang recipe for Beam

The full reference for serving Beam with SGLang.

vLLM recipe for Beam

The full reference for serving Beam with vLLM.

TokenSpeed recipe for Beam

The full reference for serving Beam with LightSeek’s TokenSpeed.

​
Serve Beam on your own compute

1

Get the weights

Check the requirements and download the weights from Hugging Face. See Download the weights.
2

Start a server

Serve Beam with SGLang, vLLM, or TokenSpeed. Each guide walks through one setup on a single server. For other GPUs and precisions, use the engine recipes above.
3

Verify the deployment

Run Beam Verifier to check your deployment’s correctness and compare its performance to reference numbers.

Serve with SGLang

Serve with vLLM

Serve with TokenSpeed

Also for self-hosting:
  • Use HF Transformers: generate text, control reasoning, and call tools with Beam in Hugging Face Transformers.
  • Beam Harmony format: how Beam’s chat template formats conversations and how its response template parses output.
  • Speculative decoding: speed up Beam’s generation with the DFlash2 drafter.
  • Quantization details: how the FP8 and NVFP4 weights were quantized.

​
Serve Beam in production

​
Use Beam through a partner

Cloud providers

Use Beam through your cloud provider.

Hardware partners

Run Beam on partner hardware and systems.

Inference providers

Call Beam through third-party inference providers.
Last modified on October 16, 2026