> ## Documentation Index
> Fetch the complete documentation index at: https://developers.reflection.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Download the weights

> Check the requirements and download Beam's weights before you serve it yourself

Download Beam's weights before you serve it with [SGLang](/serving-sglang), [vLLM](/serving-vllm), or [TokenSpeed](/serving-tokenspeed). For more about the model, see [Beam's model card](https://huggingface.co/reflection/Beam-1.0-501B-A23B).

## Requirements

<CardGroup cols={3}>
  <Card title="GPUs" icon="microchip">
    Four NVIDIA B300 SXM GPUs connected by NVLink, on a Linux amd64 host.
  </Card>

  <Card title="Docker" icon="docker" iconType="brands">
    Docker with the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html) and a compatible NVIDIA driver.
  </Card>

  <Card title="Model access" icon="key">
    Access to the Beam NVFP4 model and the DFlash2 drafter.
  </Card>
</CardGroup>

## Download the model

<Steps>
  <Step title="Install the Hugging Face CLI">
    Install the [Hugging Face CLI](https://huggingface.co/docs/huggingface_hub/en/guides/cli#getting-started), and run `hf auth login` if the model requires authentication.
  </Step>

  <Step title="Choose a download directory">
    The serving guides mount this directory into the container.

    ```bash icon="terminal" theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
    export MODEL_DIR="$HOME/beam-models"
    ```
  </Step>

  <Step title="Download the model">
    The SGLang, vLLM, and TokenSpeed guides use NVFP4. Beam is also published in BF16 and FP8.

    <Warning>
      The checkpoints are large. Check that the disk holding `MODEL_DIR` has room for the precision you download.
    </Warning>

    <Tabs>
      <Tab title="NVFP4">
        The [NVFP4 model](https://huggingface.co/reflection/Beam-1.0-501B-A23B-NVFP4) is 273 GiB.

        ```bash Download the NVFP4 model icon="terminal" theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
        hf download reflection/Beam-1.0-501B-A23B-NVFP4 \
          --local-dir "$MODEL_DIR/nvfp4"
        ```
      </Tab>

      <Tab title="FP8">
        The [FP8 model](https://huggingface.co/reflection/Beam-1.0-501B-A23B-FP8) is 468.72 GiB.

        ```bash Download the FP8 model icon="terminal" theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
        hf download reflection/Beam-1.0-501B-A23B-FP8 \
          --local-dir "$MODEL_DIR/fp8"
        ```
      </Tab>

      <Tab title="BF16">
        The [BF16 model](https://huggingface.co/reflection/Beam-1.0-501B-A23B) is 933.81 GiB.

        ```bash Download the BF16 model icon="terminal" theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
        hf download reflection/Beam-1.0-501B-A23B \
          --local-dir "$MODEL_DIR/bf16"
        ```
      </Tab>
    </Tabs>
  </Step>

  <Step title="Download the DFlash2 drafter">
    The serving guides use the [DFlash2 drafter](https://huggingface.co/reflection/Beam-1.0-501B-A23B-DFlash2) for [speculative decoding](/speculative-decoding). It works with all three precisions.

    ```bash Download the DFlash2 drafter icon="terminal" theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
    hf download reflection/Beam-1.0-501B-A23B-DFlash2 \
      --local-dir "$MODEL_DIR/dflash2"
    ```
  </Step>
</Steps>