For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

System requirements

To install MAX and serve models, your system must meet the following requirements.

For the GPUs you can program with Mojo, along with their architecture targets, drivers, and troubleshooting, see Hardware compatibility.

Linux

Requirement	Details
Operating system	Linux with glibc 2.34 or later. We test on Ubuntu 22.04 LTS and later.
CPU	x86-64 CPUs from about 2013 onward, or ARM64 Neoverse N1 or newer, such as AWS Graviton2
Memory	8 GiB minimum
Compiler	A C compiler, such as cc, gcc, or clang. MAX uses it as a linker.

macOS

Requirement	Details
Operating system	macOS Sequoia (15) or later
CPU	Apple silicon
Memory	8 GiB minimum
Developer tools	Xcode or Xcode Command Line Tools 16 or later

For GPU programming, you might also need the Metal toolchain. To install it, run the following command:

xcodebuild -downloadComponent MetalToolchain

Windows

MAX doesn't officially support Windows. To try MAX on Windows, install Windows Subsystem for Linux (WSL) with a version of Ubuntu that meets the requirements on the Linux tab.

GPU compatibility

MAX supports a wide range of GPUs for accelerated inference, and we regularly verify model compatibility on NVIDIA B200 and AMD MI355X GPUs.

For specific GPU architectures, driver versions, and troubleshooting, see Hardware compatibility.

Apple silicon GPUs

We're actively developing Apple silicon GPU support, both for GPU programming with Mojo and for model inference with MAX. Apple silicon GPUs run a subset of models and don't yet have the same CI coverage as B200 and MI355X.

For supported chips and Metal troubleshooting, see Apple silicon GPUs.

Additional considerations

• Models vary in memory requirements. Because of this, even if we list your GPU architecture as compatible, you should confirm that the available memory is sufficient for the model you're using. To check how much memory a model needs, see its Hugging Face page.

• On systems where the GPU shares system RAM (Apple silicon and integrated GPUs such as those in DGX Spark and Strix Halo), a model's weights and activations have to fit in memory alongside everything else the system is running.

• MAX can serve models on a CPU or GPU, but most of the latest models require one or more GPUs. When you browse our supported models, you can check for multi-GPU support. To serve a model across multiple GPUs, see Parallelism.

• On x86-64 hosts, your CPU must support the x86-64-v3 microarchitecture level, which requires the AVX, AVX2, BMI1, BMI2, F16C, FMA, LZCNT, MOVBE, and XSAVE instructions (this doesn't apply to ARM64 hosts). To check your CPU, run the following command and confirm that the output includes those flags:

  grep flags /proc/cpuinfo