Modular details its open AI compiler stack at LLVM meeting

Modular presented its end-to-end Mojo and MAX software stack at the 2026 LLVM Developers' Meeting, including a full-day workshop and a technical talk on device graphs. The company also outlined plans for Mojo, its compiler backends and developer tooling.
- Modular presented its full Mojo and MAX stack at the 2026 LLVM Developers' Meeting.
- The company described Mojo's MLIR-based compiler and plugin architecture.
- MAX builds device graphs directly in the compiler rather than recording them at runtime.
- Modular said direct graph construction is about 40% faster than recording.
- The public modular/modular repository has over 30,000 GitHub stars.
A deeper look at Mojo and MAX
Modular spent three days at the event in Santa Clara, California, which ran from October 26 to 28. The company said this was the first time it could share details about the entire Modular stack with all of the source code available for attendees to follow along.
Its October 26 workshop, titled Mojo: Internals Deep Dive and Roadmap, targeted compiler engineers familiar with LLVM and MLIR. The session covered Mojo's compiler, type system, GPU backends and standard library, while also showing how MAX targets accelerator hardware and compiles and optimizes computational graphs for AI models.
According to Modular, Mojo is built on MLIR, a subproject of LLVM. Presenters described it as a user-facing layer over MLIR for accelerators, with hardware specialization provided through libraries that domain experts can write. The company also discussed a plugin system intended to add new backends without forking the entire compiler. Modular said the backend for Qualcomm Cloud AI 100 is 263 lines.
MAX builds device graphs in the compiler
In a technical talk titled Moving Beyond Record & Replay, Spenser Bauman of Modular's Graph Execution team argued for constructing device graphs in a compiler rather than capturing them at runtime.
Device graphs submit a sequence of work to an accelerator at once. Modular said this can help address the growing cost of CPU-to-accelerator communication as GPUs become faster. The company described record and replay as an approach that spreads complexity across components including the graph executor, asynchronous runtime, memory pools, kernels and drivers. It also said some accelerators provide device-graph or command-buffer APIs without offering an equivalent capture mode.
MAX's graph compiler can instead construct a device graph and emit direct calls to the target's graph-construction APIs. Modular said the approach is already used by a production Gemma 4 model running end to end through device graph synthesis on one or more GPUs. According to the company, building the graph directly is about 40% faster than recording it. DeviceGraphBuilder implementations exist for CUDA, HIP and Qualcomm AI100 behind one interface.
Roadmap and open-source work
Modular's near-term Mojo priorities include enumerations and pattern matching, more compile-time metaprogramming, and syntax for hardware concepts such as address spaces. The company's longer-term goal for 2027 is to reach developers writing performant CPU code, with planned work on async/await, dynamic dispatch, threading and networking.
On the compiler side, Modular plans to make GPU and NPU backends source available, support pluggable MLIR and LLVM backends, enable a WASM backend and continue improving compilation speed. Proposed tooling includes automatic Mojo and C interoperability, a debugger, a testing framework, a package manager built on pixi and a Bazel-based build system.
The company said its open-source work began with the Mojo standard library in March of 2024 and progressed through open-sourcing Mojo and making MAX fully source-available a few weeks before the event. Modular said the entire stack now builds from the public modular/modular repository, which has over 30,000 GitHub stars. For developers and hardware partners, the presentations offer a view of how Modular intends to use Mojo, MLIR, plugins and MAX to extend its software stack across accelerator targets.
Modular: Modular: Modular at LLVM Developers' Meeting 2026


