modular/max-full repository overview

⁠MAX full container

The Modular⁠ platform, powered by the MAX framework and Mojo programming language, helps developers accelerate model serving and scale GenAI deployments across diverse hardware architectures without modifying a single line of code.

MAX delivers high-performance, hardware-agnostic inference for popular open models, with seamless portability across cloud providers and devices.

The max-full container includes all the core dependencies needed to deploy LLMs on both NVIDIA and AMD GPUs. It offers a fully integrated environment with support for MAX models, PyTorch (GPU), ROCm, CUDA, and cuDNN, ensuring optimal performance across GPU types. Use this container if you want a streamlined, out-of-the-box solution that works across GPU vendors without additional setup.

The MAX container is compatible with the OpenAI API specification and optimized for GPU deployment. For details on container contents and hardware compatibility, see MAX containers⁠ in the MAX documentation.

⁠Quickstart

You can run an LLM on GPU using the latest MAX full container with the following commands.

Run the container on an AMD GPU:

docker run \
  # HuggingFace configs
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --env "HF_TOKEN=$HF_TOKEN" \
  # AMD configs
  --group-add keep-groups \
  --rm \
  --device /dev/kfd \
  --device /dev/dri \
  # MAX configs
  -p 8000:8000 \
  modular/max-full:<version> \
  --model-path <model-provider/model-id>

For more information on AMD-specific command configurations, see Running ROCm Docker containers⁠.

Run the container on an NVIDIA GPU:

docker run \
  --gpus 1 \
  # HuggingFace configs
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --env "HF_TOKEN=<secret>" \
  # MAX configs
  -p 8000:8000 \
  modular/max-full:<version> \
  --model-path <model-provider/model-id>

You can run a MAX model⁠ by referencing its HuggingFace model ID. For example, google/gemma-3-1b-it⁠.

You can also use the MAX container to run a variety of PyTorch models⁠ hosted on Hugging Face, such as microsoft/Phi-3.5-vision-instruct⁠.

For more information on deploying popular models with MAX, see the model support⁠ documentation.

⁠Tags

Supported tags are updated to the latest MAX versions, which include the latest stable release and more experimental nightly releases. The latest tag provides you with the latest stable version and the nightly tag provides you with the latest nightly version.

Stable

• max-full:25.X

Nightlies

• max-full:25.X.0.devYYYYMMDD

⁠Documentation

For more information on Modular and its products, visit the Modular documentation site⁠⁠⁠.

⁠Community

To stay up to date with new releases, sign up for our newsletter⁠⁠⁠, check out the community⁠⁠⁠, and join our forum⁠⁠⁠.

If you're interested in becoming a design partner to get early access and give us feedback, please contact us⁠⁠⁠.