Modular MAX, the company's stack for building and serving generative AI models, is now fully source available. The graph compiler, the runtime and the hardware backends for CPUs and GPUs were the core pieces still closed, and all three can now be read, built and modified. The license that covers them forbids feeding any of it to an AI system to build a replacement.
The Modular MAX release finishes the staged opening Modular began with the Mojo standard library in 2024. MAX's kernels, models and serving code followed through 2025, and the Mojo compiler in August. Modular says the MAX components already open, among them the kernel library, the models and the Python APIs, will remain under Apache 2.0, while the new MAX License covers the internals opened now plus the MAX binaries Modular distributes. Each source file carries a header naming its license. MAX can be used on any hardware and in commercial products without fees or royalties, according to the company.
| Component | License |
|---|---|
| Already-open MAX components (kernels, models, Python APIs and more) | Apache 2.0 |
| Graph compiler, runtime, CPU and GPU backends | MAX License |
| MAX binaries Modular distributes | MAX License |
Developers can license their own libraries and applications built on Modular MAX however they choose. Derivative works of MAX itself may be distributed only in object code form, and modified MAX source may be published only in a fork meant to contribute changes back to the main project, according to the MAX License. A company that sells hosted inference or training on MAX must use the "Modular Accelerated" word mark or logo on that service and get Modular's written approval for the materials that carry it. The license ties that duty to legal entities that deploy MAX commercially for other legal entities.
Modular's summary of the AI clause says no part of MAX may be used as input to an AI system to produce a reimplementation. The license text sets the scope. It bars using Modular MAX as training data, fine-tuning data or input to produce software that reimplements or substitutes for MAX, with or without literal copies. It also says the clause does not restrict using AI tools to read, analyze, improve or explain MAX, or to write software that runs on or interoperates with it. On that wording, asking an assistant to explain the graph compiler would fall outside the ban, while asking one to write a drop-in replacement from the source would fall inside it.
The license also lets Modular and its affiliates use MAX telemetry and usage data for any purpose and states that, as between Modular and the user, Modular owns that data. Personal information is handled under Modular's privacy policy, the company says.
Modular's explanation for not choosing Apache 2.0 is fragmentation. In the release notes for Modular 26.7, the version that ships the source, the company says the new license allows nearly everything Apache 2.0 does, and that incompatible platform-specific versions would undermine MAX's aim of one programming model across hardware. The same version adds an eager mode that shares a code path with the graph compiler. On a slice of Gemma 4 31B, Modular reports the first token at about 30 seconds in eager mode against about 110 seconds compiled.
Contributions open with bug fixes, with refactors and other changes to follow over time. Modular plans to release source for MAX plugins for other chips, including Qualcomm Dragonfly and Google TPU, in the coming months, and cofounders Chris Lattner and Tim Davis will hold a live Q&A on the license on November 10 at 10 AM PT.
