Modular open-sources Mojo three weeks after Qualcomm acquisition

Why it matters

Mojo gives Qualcomm-owned Modular an open developer layer for its cross-chip AI stack. Its credibility now depends on compiler governance and continued support for rival hardware.

Chris Lattner (@clattner_llvm) and Tim Davis (@iamtimdavis) founded Modular, which open-sourced the entire Mojo programming language on August 18, including its compiler and developer tooling, three weeks after Qualcomm completed its acquisition of Modular.

The release puts an Apache 2.0 license on the central developer project behind the founders' four-and-a-half-year effort to build a common software layer for AI hardware. It also gives Lattner and Davis an early test of the promise they carried into Qualcomm: Modular can belong to a chipmaker while continuing to support chips made by its owner's competitors.

Modular disclosed the release in its ModCon announcement, one week after Mojo reached version 1.0 on August 11. The 1.0 release established a more stable language contract after years of frequent changes, while the open-source release moved the compiler into Modular's public repository alongside the standard library and other platform components.

Few founders are better suited to make a programming language the front door to an infrastructure business. Lattner created LLVM, helped create Clang, drove the development of Apple's Swift language and later helped create Google's MLIR compiler framework. His work has repeatedly followed the same pattern: establish a lower-level technical foundation, open enough of it to attract contributors, then let other products and platforms build above it.

Davis brought the AI deployment side of that thesis. Before co-founding Modular in 2022, he worked at Google on TensorFlow, MLIR, TPU infrastructure and machine-learning systems for devices. Lattner and Davis met at Google, where the growing collection of separate software stacks for CPUs, GPUs and custom accelerators convinced them that AI infrastructure would need a common layer spanning hardware vendors.

The compiler is public

Mojo is a systems-oriented language designed for high-performance software across CPUs, GPUs and other accelerators. Its syntax and Python interoperability are meant to give AI developers a gradual path from familiar Python programs into lower-level kernels and systems code without maintaining a separate CUDA or C++ codebase for every performance-critical operation.

The public Modular repository now includes the Mojo compiler, standard library, examples and other platform components. The Mojo language, compiler and developer tooling are available under Apache 2.0.

That openness matters because Lattner has spent much of his career building projects whose influence came from shared governance and broad industrial participation. Publishing source code creates the legal base for that participation. The treatment of outside compiler changes will shape whether Mojo develops shared governance or remains a Qualcomm-owned codebase that the public can fork.

The opening stops at a crucial commercial boundary. MAX, Modular's inference framework and serving layer, is not covered by Mojo's Apache 2.0 release. Modular said at ModCon that the MAX license no longer contains device-use restrictions, that MAX will be source-available and that it is developing an open alliance program. The final scope of that program remains unclear.

Qualcomm inherited a neutrality problem

Qualcomm completed its purchase of Modular on July 29 and appointed Lattner executive vice president of advanced AI software and platforms. Mojo, MAX and Modular Cloud retained their existing brands.

RuntimeWire reported when the deal was announced that Qualcomm was buying its way into the AI runtime layer, where software determines which chips developers can use and how easily workloads move between them. That strategy gives Qualcomm a route into data centers and developer workflows that extends beyond selling its own processors.

The ownership creates an obvious tension. Modular's pitch depends on developers and hardware companies believing that MAX and Mojo will treat competing silicon as first-class targets. Qualcomm benefits when workloads run on Qualcomm hardware, while Modular gains relevance by making the underlying chip interchangeable.

Lattner and Davis addressed that tension directly at ModCon. Modular said its platform would continue to support hardware that competes with Qualcomm products. It announced support for AWS Trainium, Google TPUs, Qualcomm Cloud AI 100 Ultra and Qualcomm Dragonfly, alongside the NVIDIA and AMD GPU deployments it already offers. Those new accelerator integrations are scheduled to move into production over the coming months.

Open-sourcing Mojo gives that claim more weight because hardware vendors can inspect and adapt the language layer themselves. HTEC engineers, for example, brought Google TPU support to Modular's platform with Modular serving in a supporting role, according to the company. Modular claims the new hardware integrations required a more than 10x reduction in engineering effort compared with its traditional bring-up efforts; that comparison has not been independently benchmarked.

The language feeds a cloud business

The same keynote made Modular Cloud generally available, turning the founders' compiler and inference work into a managed service. Modular offers shared, OpenAI-compatible endpoints priced by token and dedicated deployments running on reserved infrastructure operated by Modular or its customers.

Modular identified MiniMax as a flagship customer running the M3 model on a dedicated deployment. Modular says that deployment serves billions of tokens per minute and required specialized kernels for MiniMax Sparse Attention, a technique intended to reduce the work required as context length grows. The throughput figure is company-supplied and has not been independently verified.

This explains why Modular chose to open Mojo while retaining different terms around MAX. A permissively licensed language can expand the developer base, attract hardware ports and make Mojo a default way to write specialized kernels. MAX and Modular Cloud then capture commercial value from deploying and serving the resulting models.

The founders had substantial financial backing to build that stack before the Qualcomm sale. Modular announced a $250 million round on September 24, 2025, led by U.S. Innovative Technology Fund, with DFJ Growth, GV, General Catalyst and Greylock participating. Modular said the financing valued it at $1.6 billion and brought its total capital raised to $380 million.

Qualcomm now supplies the balance sheet and hardware reach that Modular once needed venture capital to pursue. Open-source Mojo is Lattner and Davis's mechanism for keeping the developer side of the company credible inside that larger corporate structure.

The source release settles one long-running question around Mojo: developers can finally inspect and build on the compiler without waiting for Modular's next binary release. Governance, outside contributions and continued investment in rival hardware will determine how far the language can spread beyond its corporate owner.