Modular open sourced Mojo, its Python-like programming language for AI workloads, releasing the compiler and toolchain under an Apache 2.0 license, the company said in a blog post published at its ModCon conference.
The move follows through on a promise Modular made when it first previewed Mojo in 2023, and comes about a week after the language reached its 1.0 release, according to developer Simon Willison, who covered the announcement. Mojo was originally pitched as a superset of Python that could reuse existing Python code and libraries, though that plan has since changed, Willison wrote.
Modular also said its MAX inference software is moving to a source available license with no device restrictions, and that Modular Cloud is now publicly available at console.modular.com. The company added support for AWS Trainium, Google TPUs and two Qualcomm accelerator lines, and said native Windows support for Mojo is coming through a partnership with Microsoft, according to the announcement.
Modular said its platform now serves billions of tokens per minute in production, including a dedicated deployment running MiniMax's M3 model, and that its heterogeneous hardware support required more than 10 times less engineering effort than its earlier approach, according to the blog post.
An Apache 2.0 compiler means anyone can now fork, audit or extend Mojo without waiting on Modular, which matters if you held off building on the language until the tooling itself was open. Multi-vendor accelerator support is the other half of the bet: if it holds up, it is one less reason to write kernels twice for two different chip vendors.