AI Potluck
Contribute
Back to Gap Map Model components / Inference code

MAX

Modular
source available / Overall score: 3.5

MAX is an inference engine built on the Mojo programming language that serves LLMs and other models across NVIDIA, AMD, and Apple GPUs plus CPUs from a single codebase. A graph compiler optimizes model execution for the target hardware, and it serves PyTorch and Hugging Face models through OpenAI-compatible endpoints. Modular, the company founded by LLVM and Swift creator Chris Lattner, develops and distributes it.

The repository source and MAX/Mojo usage carry different licenses; the openness score follows the usage terms rather than the repository file.

Openness

2 high confidence
2.0
license
Modular-Community-License(non-OSI
repo-license
Apache-2.0-WITH-LLVM-exception(the source in modular/modular)
source
public(~450k LOC incl MAX kernels + inference server + Mojo stdlib)
commercial-tier
Modular enterprise/managed deployment

MAX's own repository code is Apache-2.0, but using or distributing MAX itself falls under a separate, non-OSI Modular Community License. That license limits redistribution to object code and restricts other uses, so the source is freely runnable but not unconditionally open.

  • https://github.com/modular/modular recorded 2026-09-09

    MAX usage and distribution are licensed under the Modular Community License, and the repository's own code is licensed under the Apache License v2.0 with LLVM Exceptions. The line recorded on 2026-08-09 named Modular and Mojo alongside MAX; it now names MAX only, and points at the same license.

  • https://raw.githubusercontent.com/modular/modular/main/LICENSE recorded 2026-10-14

    The MAX repository is licensed under the Apache License v2.0 with LLVM Exceptions:

  • https://www.modular.com/legal/community recorded 2026-10-14

    The old license capped free production use at eight accelerators outside x86, ARM and NVIDIA. Both requirements are now gone. You shall not use MAX, or any portion of it, as training data, fine-tuning data, or input to any AI system in order to produce software that reimplements or substitutes for MAX. Last Modified: August 18, 2026.

Adoption

3 low confidence
3.0

MAX serving publishes no download, user or token figure, leaving GitHub stars as the only available signal. A star is not a download, and modular/modular is a single repository shared by MAX and Mojo, so an unknown share of its stars belongs to the language rather than to the serving product.

Capability

4 medium confidence
4.0

MAX serves hundreds of open models, including DeepSeek, Gemma and Qwen, across NVIDIA, AMD and Apple GPUs and CPUs through an OpenAI-compatible API, with custom kernel authoring in Mojo. The speed comparison in its documentation is Modular's own, against vLLM, and no MLPerf or Artificial Analysis result is cited, so no independent benchmark supports its performance.

  • https://docs.modular.com/max/intro recorded 2026-09-29

    The intro page says MAX serves open models through an OpenAI-compatible endpoint: "Deploy DeepSeek, Gemma, Qwen, and hundreds more with an OpenAI-compatible endpoint on NVIDIA or AMD and on any cloud." It also says "The same MAX package runs on NVIDIA, AMD, and Apple silicon". The earlier 500+ model count is no longer stated on the page.

  • https://docs.modular.com/max/models recorded 2026-09-29

    The supported-models page lists architectures: "The table below lists all the model architectures currently supported by MAX."

  • https://github.com/modular/modular recorded 2026-08-09

    We have separate containers for NVIDIA and AMD GPU environments, and a unified container that works with both.

Verified 2026-09-09