Modular Platform Files

In this release, we've moved most of the GPU programming APIs out of the Mojo standard library and into a new top-level max package—std.algorithm is now max.algorithm, std.benchmark is max.benchmark, and std.gpu.{compute,host,memory,sync} is max.gpu.*, along with the Mojo layout package. All of which is now part of the new MAX accelerator library, which also includes the existing AI kernels. We've also extended Apple silicon GPU support back to M1 devices, and added hardware-MMA flash-attention prefill for M5 to speed up TTFT. The MAX model bring-up workflow also gains new agent skills in our skills repo, including new serve-model, benchmark-model, eval-model, and profile-model skills. We've also expanded model coverage and capabilities, made significant strides speeding up the Python API's experimental eager interpreter, sped up various kernels across NVIDIA, AMD, and Apple GPUs, and much more to accelerate models served with MAX.

Of course, this release also brings Mojo 1.0! 🔥

With Mojo 1.0, we've started marking standard library APIs as stable, which means they won't be removed or changed in ways that break compatibility—we started with a small set that we'll grow in subsequent 1.x releases. There are so many changes in this release, it's difficult to summarize. It includes lambdas, pointer unification, interior origins, required var declarations, constraint improvements, faster Python interop, and so much more. We worked hard on getting names, defaults, and safety boundaries right for 1.0. That does mean you'll find more breaking changes than usual in this release. But nearly every breaking change ships with a deprecated alias and a compiler fix-it, so migration is mechanical.

Check out all the updates with the full MAX release notes and Mojo release notes.