Modular makes the full MAX inference stack source available alongside Apache 2.0 Mojo 1.2
I'm happy to share that Mojo and MAX are now fully open - for development, extension, and commercial use! This includes the full Mojo programming language, but also the MAX graph compiler, MLRT runtime internals, and hardware abstraction logic. Get your agents going!👇
Context
Opening the MAX graph compiler, MLRT runtime internals, and hardware abstraction layer lets developers and AI coding agents inspect and extend the entire compiler stack across heterogeneous CPUs and GPUs without licensing fees.
Modular open sourced the Mojo standard library in 2024 and the Mojo 1.0 compiler under the Apache 2.0 license with LLVM exceptions in August 2026 after entering Qualcomm. Until the MAX 26.7 release, the core MAX graph compiler and runtime internals remained closed source even as MAX kernels and serving APIs were opened.
Deep dive
Modular cofounder Chris Lattner announced on October 23 that Mojo and MAX are now open for development, extension, and commercial use with the release of MAX 26.7 and Mojo 1.2. While the Mojo programming language and MAX kernel libraries remain under the Apache 2.0 license with LLVM exceptions, Modular published the remaining core components of its inference stack under a new source available MAX License. Those newly opened components include the MAX graph compiler, the MLRT runtime internals, and the hardware abstraction logic for CPUs and GPUs.
In its release blog post, Modular explained that the MAX License permits unlimited commercial deployment on any hardware without licensing fees or royalties, and allows developers to fork the repository to contribute bug fixes and improvements back upstream. However, the MAX License includes three notable restrictions that distinguish it from standard open source licenses: users cannot feed MAX source code into AI systems to generate a competing reimplementation, commercial hosted inference or training services built on MAX must display a Modular Accelerated wordmark, and developers cannot publish or distribute modified MAX source code outside of upstream contributions.
Alongside the license change, MAX 26.7 adds support for Muse Glimmer and MiMo V2.6 Flash, video input for Qwen3.5 and Qwen3 VL, a Decisions API for classification probabilities, Jenga as the default KV cache, and performance gains on Nvidia B200 and AMD MI355 GPUs. Mojo 1.2 adds structural pattern matching and shared CPU and GPU math functions, while Modular prepares a broader hardware and model lab alliance program.
What to expect next
Making the MLRT runtime and graph compiler source available gives developers and coding agents direct visibility into how high level model graphs lower to custom accelerator kernels across CPUs and GPUs. Whether third party chipmakers and enterprise platform teams adopt MAX as a shared alternative to proprietary vendor stacks will depend on how Modular structures governance, upstream review, and fork permissions in its upcoming alliance program.