Download the weights

Download Beam’s weights before you serve it with SGLang, vLLM, or TokenSpeed. For more about the model, see Beam’s model card.

Requirements

Four NVIDIA B300 SXM GPUs connected by NVLink, on a Linux amd64 host.

Docker with the NVIDIA Container Toolkit and a compatible NVIDIA driver.

Access to the Beam NVFP4 model and the DFlash2 drafter.

Install the Hugging Face CLI

Install the Hugging Face CLI, and run hf auth login if the model requires authentication.

Choose a download directory

The serving guides mount this directory into the container.

export MODEL_DIR="$HOME/beam-models"

Download the model

The SGLang, vLLM, and TokenSpeed guides use NVFP4. Beam is also published in BF16 and FP8.

The checkpoints are large. Check that the disk holding MODEL_DIR has room for the precision you download.

The NVFP4 model is 273 GiB.

Download the NVFP4 model

hf download reflection/Beam-1.0-501B-A23B-NVFP4 \
  --local-dir "$MODEL_DIR/nvfp4"

The FP8 model is 468.72 GiB.

Download the FP8 model

hf download reflection/Beam-1.0-501B-A23B-FP8 \
  --local-dir "$MODEL_DIR/fp8"

The BF16 model is 933.81 GiB.

Download the BF16 model

hf download reflection/Beam-1.0-501B-A23B \
  --local-dir "$MODEL_DIR/bf16"

Download the DFlash2 drafter

The serving guides use the DFlash2 drafter for speculative decoding. It works with all three precisions.

Download the DFlash2 drafter

hf download reflection/Beam-1.0-501B-A23B-DFlash2 \
  --local-dir "$MODEL_DIR/dflash2"

Last modified on October 16, 2026