IMPORTANT: To view this page as Markdown, append `.md` to the URL (e.g. /get-started.md). For the complete documentation index, see llms.txt.
Skip to main content
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).

Packages

You can install all the MAX API libraries and tools, including Mojo and all optional dependencies, with max[all] (pip/uv) or max-all (conda/pixi). Alternatively, you can install just a subset of the tools using one of the other max package options as described below.

Install​

To get the latest improvements and new features, we recommend installing our nightly build, which we release several times a week. If you want a better tested but older version, you can install a stable build. For version details, see the release notes.

System requirements:

  1. If you don't have it, install pixi:
    curl -fsSL https://pixi.sh/install.sh | sh

    Then restart your terminal for the changes to take effect.

  2. Create a project:
    pixi init example-project \
  -c https://conda.modular.com/max-nightly/ -c conda-forge \
  && cd example-project
  3. Install max with all dependencies (stableTo get the nightly build, change the version in the website header.):
    pixi add "max-all==26.7"
  4. Start the virtual environment:
    pixi shell

See the following section for more details about what's included in max[all] (and max-all).

What's included​

The max[all] (or max-all) package includes all optional dependencies so you can do everything, including serve models, benchmark endpoints, and program with Mojo.

If you're concerned about bloat in your dependencies, you can chose a smaller package that includes only what you need. For example, if you only need to benchmark an endpoint, you can install just max[benchmark].

max[all]​

  • pip/uv: max[all]
  • conda/pixi: max-all

This is the "all extras" package. It includes everything MAX has to offer.

It's just a convenient way to get all these packages at once:

  • max[benchmark]
  • max[serve]
  • mojo

max[benchmark]​

  • pip/uv: max[benchmark]
  • conda/pixi: max-benchmark

This is required to use the max benchmark command.

It depends on max and other third-party packages.

max[serve]​

  • pip/uv: max[serve]
  • conda/pixi: max-serve

This is required to create an endpoint with the max serve command.

It depends on max and other third-party packages.

max​

This is the main package you need to do anything with MAX. It includes:

max depends on mojo-compiler and minimal third-party packages.

If you want all Mojo developer tools, you must install the larger mojo package, which is included in the max[all] package. To see what's in the mojo and mojo-compiler packages, see what's include in the Mojo SDK.

Technically, there are other packages you'll see in your dependency tree (such as max-core) but they're implementation detail and you should generally ignore them—don't set them as explicit dependencies in your project because they won't work on their own.

System requirements​

Your system must meet these specifications.

  • macOS Sequoia (15) or later.
  • Apple silicon (M1–M5 processor).
  • Python 3.10–3.14.
  • Xcode or Xcode Command Line Tools 16 or later.
  • Apple silicon GPU support is available for Mojo GPU programming. See GPU compatibility for details. You may need to run xcodebuild -downloadComponent MetalToolchain to install the Metal utilities required for GPU programming.

GPU compatibility​

MAX supports both CPUs and GPUs, so you don't always need a GPU to serve a model. To accelerate inference with GPUs, we test serving on NVIDIA B200 and AMD MI355X GPUs. Other compatible GPUs might also serve models, but we don't test them for serving. For every GPU you can program with Mojo and MAX, and the arch target of each, see Hardware compatibility.

GPUs require the following drivers:

  • NVIDIA: GPU driver 580 or later. Check your version with nvidia-smi. To update, see the NVIDIA driver docs. If your driver is older than 580, set the MODULAR_NVPTX_COMPILER_PATH environment variable to a system ptxas binary (for example, export MODULAR_NVPTX_COMPILER_PATH=/usr/local/cuda/bin/ptxas).
  • AMD: GPU driver 6.3.3 or later (MI355X requires ROCm 7.0 or later). For installation, see the Ubuntu native install guide.

Apple silicon GPUs​

We're actively developing Apple silicon GPU support (M1–M5), both for GPU programming with Mojo and for model inference with MAX.

MAX supports a subset of models on Apple silicon today. We've manually run the Llama, Gemma, Nemotron, and FLUX.2 families, and while Apple silicon isn't yet covered by the same nightly CI as the GPUs listed above, support continues to mature.

Quantization moves the memory ceiling significantly; for example, FLUX.2 dev needs over 120 GB at BF16 and roughly 80 GB at FP4. The FP4 kernels target M5, and support for other generations is ongoing.

Architectures we haven't exercised on Metal may depend on kernels that don't exist yet. When that happens, the model fails during graph compilation rather than at inference time.

For hardware tables and Metal troubleshooting, see Apple silicon GPUs.

Edit this page

Was this page helpful?

Models

Cookbook

Puzzles

Blog

Community