All articles
Dev ToolsBy Petru Popa · 5 min read

The Mojo License Is Apache. The MAX SDK License Is Not.

Modular open sourced the Mojo compiler on 18 August. The repository holds two licenses, and the announcement names one of them.

Modular put the Mojo compiler on GitHub on 18 August under the Apache License v2.0 with LLVM exceptions. It is not in the directory named mojo — that one holds the standard library, documentation, proposals, examples and Python integration. The compiler is in KGEN, a C++ tree of parser, MLIR dialects, optimization passes and a debugger, whose name the repository documentation explains as kernel generator.

What the repository offers a team writing accelerator code is one source language that compiles for NVIDIA, AMD and Apple GPUs, plus the kernel library that ships in the same repository as the MAX inference server. The second part is what you would otherwise hand-roll: a kernel set maintained against three vendors, in public, by the people shipping it.

Two questions decide whether you can depend on that. Which license governs the artifact you redistribute, and who fixes the compiler bug that blocks your release. Both are answerable from the repository. Neither is answered by the announcement.

What the Mojo license actually covers

The root LICENSE file is the Apache License v2.0 with the LLVM exceptions, and the per-file headers agree with it. max/mojo/max/__init__.mojo — the entry point of the package that the 26.5 release moved the GPU APIs into, when std.gpu.* became max.gpu.* — carries the Apache header and a 2026 Modular copyright. The source in the repository reads as Apache: compiler, standard library, kernels, serving code.

The README then adds a sentence the announcement does not. MAX usage and distribution are licensed under the Modular Community License. That license is in the repository too, at Licenses/LICENSE, and it carries a last-modified date of 18 August 2026 — the day of the announcement. Its terms are not the Apache terms. Derivative works may be distributed in object code form only. An application you distribute must carry material additional functionality beyond the included portions of MAX. You must display a notice naming Modular. The patent grant ends if you litigate against Modular. And you may not use MAX as training data, fine-tuning data or input to an AI system in order to produce software that reimplements or substitutes for it.

The gap is that this license never names a package, a wheel or a file. It defines its subject as the MAX software development kits and stops there. Licenses/README.md says the directory holds license files for external Modular product packages and that licenses may differ per product, without saying which product gets which. The repository tells you two licenses exist and does not draw the boundary between them.

GitHub's classifier does not draw it either. The license endpoint for the repository returns a key of other, a name of Other, and an SPDX identifier of NOASSERTION: GitHub cannot resolve this repository to a recognized license. The SPDX license list carries Apache-2.0, marked both OSI approved and FSF free. It carries no Modular entry.

The maturity signal is the pull request queue

Activity is not in doubt. 1,815 commits landed on main in the thirty days to 19 August, across 39 distinct author names in a hundred-commit sample of that window. Five releases shipped in 2026 — 29 January, 19 March, 7 May, 18 June and 11 August — a six to eight week cadence. The repository was created in April 2023 and has 27,399 stars.

The pull request numbers are a different shape. Across 2026, nine pull requests have been merged and 412 were closed without merging. 295 are open, 49 of them opened more than a year ago, and the oldest is a standard library change, not a draft, open since 28 June 2024. The standard library has taken outside contributions since 2024. The README says compiler contributions are not accepted yet, and Modular's announcement states an aim to accept them by the end of the year.

Nine merges against 1,815 commits in a month is the number to plan around. Code lands here by internal push. The pull request queue is a suggestion box, not the merge path. Reading the compiler is now permitted; changing it, on the evidence in the repository, is not yet something outsiders do.

The issue tracker is where that bites. Issues filed on 18 and 19 August include a compiler hang on an Optional[SIMD[DType.bool, 2]], a crash with no diagnostic at all, and a GPU kernel that silently loses its writes when a heap allocation happens inside the kernel — no error, no warning. Those are the bugs that cost a release. 213 issues opened before August 2024 are still open.

What to do this quarter

The license question is answerable in an afternoon, and it is the one to answer before the benchmark work starts. Three checks:

  • Run gh api repos/modular/modular/license and read the NOASSERTION yourself, before a legal review reads a summary that says Apache.
  • List the artifacts you would actually ship — wheels, containers, static libraries — and for each one find the license text inside the package, not in the repository. The repository being Apache does not tell you what a MAX SDK package carries.
  • If any shipped artifact includes MAX, test your product against the material additional functionality condition and the attribution notice. A thin serving wrapper is the case that condition describes.

The claim here is falsifiable on a date. If by 31 December Modular is merging outside pull requests at a rate comparable to its internal commit volume, and has published a per-artifact license manifest, then what the repository shows today was a snapshot of an open sourcing in progress rather than its steady state. The same two commands settle it either way.

This is the cost that buy, build, or fine-tune decisions underprice. Building on a stack means inheriting its release process and its license boundary, and neither shows up in a benchmark table. Mojo is worth reading. Depending on it is a decision about a merge queue.

Ready to start?

Turn this into a plan for your team.

One week, fixed fee: a working session with your team, a prioritized use-case backlog, and an ROI model for the opportunities worth chasing.

Book an AI Opportunity Sprint