Qualcomm Open-Sources Modular AI Platform and Mojo Language

Qualcomm open-sourced Modular’s full AI software platform and the Mojo programming language at ModCon 2026. The event marked the first conference Modular held as a Qualcomm company. The announcement came on August 18 in San Francisco. The complete Mojo compiler and the broader Modular stack now sit under the Apache 2.0 license with LLVM exceptions. The code is available on GitHub.

The platform aims to challenge NVIDIA CUDA’s hold on AI software infrastructure. It lets AI inference run across Qualcomm Snapdragon chips, AMD GPUs, NVIDIA GPUs, Apple silicon, AWS Trainium, and Google TPUs inside one framework.

From Acquisition to Open Source

Qualcomm finished its roughly $3.9 billion purchase of Modular in late July. The deal had been announced on June 24. The goal was to strengthen Qualcomm’s AI software across devices, edge infrastructure, and data centers.

Instead of keeping the technology closed, Qualcomm released it as open source. ServeTheHome noted the decision stands out because Qualcomm spent billions on Modular.

AMD Joins the Effort

AMD took the stage at ModCon to show support for the Modular platform. ServeTheHome called the appearance notable. AMD earns substantial revenue from AI GPUs and could otherwise treat Qualcomm as a competitor. The platform already runs on AMD Ryzen AI Halo systems and AMD GPUs at scale.

Modular also showed the software working on a Chinese Houmo 24GB M.2 AI accelerator. That demo underscored how many different pieces of hardware the framework can target.

Mojo 1.0 Goes Fully Open

Chris Lattner, known for LLVM and Swift, created the Mojo language. It hit its 1.0 milestone on August 12. At ModCon the team delivered on long-standing promises and open-sourced the full compiler.

Mojo is built for varied hardware—CPUs, GPUs, and AI accelerators. It works with Python. Modular is adding Windows support while macOS and Linux versions already exist.

GV managing partner Dave Munichiello said Modular lets “any AI model to run on any chip.” He called it an alternative to CUDA and an example of opportunity in the AI application layer.