Modular opens up MAX: full source now available under a permissive license

MAX source opens under a permissive license

According to Modular, the full source for MAX is now available, including hardware backends, the graph compiler and the runtime. The company says hardware innovators, researchers and others can now build and modify it. It had committed to the step at its recent ModCon conference.

The new MAX License allows users to use, modify, extend and build commercial products with MAX, without licensing fees or royalties, according to Modular. Developers who want to bring up MAX on their own accelerator can use the existing implementations for NVIDIA, AMD, Apple Silicon and CPUs as a reference. Modular plans to release source for MAX plugins for other chips, including Qualcomm Dragonfly and Google TPU, in the coming months.

Modular also explains why it did not choose Apache 2.0. In its words, the MAX License allows nearly everything Apache 2.0 does. The difference is about fragmentation: MAX aims to unify the programming model across hardware, and incompatible platform-specific versions would undermine that goal. The company points to its alliance program as the way for hardware vendors and partners to build on MAX without splintering the ecosystem. It adds that, as with Mojo, it expects its approach to evolve as the technology matures and it hears from the community.

Eager mode, decision models and faster prefill

MAX now has full eager mode support, including custom ops. Modular says eager mode shares a single code path with the graph compiler, so code run in eager mode is the same as code that is compiled and shipped. A developer can declare an op once, call it directly on a Tensor, and reuse it unchanged inside a compiled Module.

The company says this shortens the development loop. On a slice of Gemma 4 31B, eager mode reaches the first token in about 30 seconds, compared with about 110 seconds when compiled.

On the model side, MAX has added initial support for System One and decision models, including Decider 2B, through the new v1/decisions and v1/systemone API endpoints. Modular says these models enable zero-shot classification at very low cost. MAX also now supports MiMo-V2.6-Flash, GLM-5.3-Flash and Muse-Glimmer, with KDA and sparse MLA.

For existing model families, several ops that used to round-trip through the CPU now run entirely on the GPU. According to Modular, this gives a 2.6x prefill speedup for Qwen and GLM models.

Mojo 1.2 adds pattern matching

Mojo 1.2 continues the language's move toward systems programming, the next phase of its roadmap after Mojo 1.0. Its flagship feature is pattern matching. The match statement replaces long chains of conditionals with structured cases, and the compiler can check that every case has been handled. Modular says it also works at compile time.

The release also makes traits and constraints more expressive. Structs can explicitly opt out of trait conformance, for example by declaring themselves not Movable. Where clauses now accept an else branch, so developers can state exactly why a constraint failed. Modular says both changes make generic code easier to read and debug.

What it means for developers and hardware vendors

For chip makers and researchers, the release removes the main barrier to experimenting with MAX on new accelerators: the code is available, and commercial use carries no fees or royalties. The alliance program is Modular's proposed route for vendors who want to contribute without creating incompatible forks.

For application developers, the eager mode offers quicker iteration on custom ops without a separate code path. Users can upgrade with pip install --upgrade max. Modular has also published a MAX License blog post and FAQ, and invites developers to share what they build on its forum.