Modular, Now Owned by Qualcomm, Fully Open-Sources Mojo Language

The announcement came during a keynote at ModCon, a developer event held in San Francisco on August 19. Modular was founded four and a half years ago on the premise that "AI will not run on just one type of chip forever," and the company presented this release as the payoff of that bet.

Three stages of openness, from standard library to the full language

Mojo's move to open source didn't happen all at once. Modular first released the standard library in 2024, then opened the kernels of its inference engine MAX in 2025. With the compiler and the rest of the language now released, the three-stage process is complete. Modular also announced that the language reached its stable 1.0 release last week. The core of 1.0 is a stability guarantee — a promise that code written today won't suddenly break with future updates.

Mojo has so far supported macOS and Linux, with Windows users required to go through WSL (Windows Subsystem for Linux). Modular said it is working with Microsoft's Windows team to bring native Windows support.

Now under Qualcomm, Modular keeps supporting rival chips

The announcement coincides with the completion of Qualcomm's acquisition of Modular. The company said that even after the acquisition, its platform will continue to support hardware that directly competes with Qualcomm. Modular explained: "A foundation only works if everyone can stand on it."

In practice, supported hardware has expanded from NVIDIA and AMD GPUs to AWS Trainium, Google TPUs, Qualcomm Cloud AI 100 Ultra, and Qualcomm Dragonfly — all running on the same modeling API, serving workflow, programming language, and core abstractions. Modular said the engineering effort required to bring new hardware onto the platform has dropped by more than 10x compared to previous approaches. Partner HTEC reportedly implemented Google TPU support on its own with just a handful of engineers over a few months.

MiniMax shows real-world performance

Modular Cloud also moved to general availability that same day at console.modular.com. It offers two options: a shared, pay-per-token endpoint compatible with the OpenAI API, and dedicated deployments that run on Modular's own infrastructure or a customer's infrastructure.

The service has been handling OpenRouter traffic under the name "ModelRun" for the past several months, and according to independent benchmarking firm Artificial Analysis, it has ranked near the top in both latency and throughput.

MiniMax stands out as a flagship example. The company runs its M3 model on a dedicated Modular Cloud deployment, handling live production traffic at a scale of billions of tokens per minute. M3 combines a 1-million-token context window, multimodal processing, and MSA (MiniMax Sparse Attention) — a sparse attention technique that reduces computation by selecting only the necessary key-value blocks. Modular said it ported the model directly onto MAX and tuned dedicated MSA kernels to optimize for the service's actual traffic patterns.

Opening up MAX and an industry alliance program

Modular is also changing the licensing policy for its inference engine MAX, expanding source access so that developers, enterprises, hardware vendors, and partners can directly extend and contribute to the platform. The company also said it is preparing an industry alliance program spanning hardware vendors, model developers, cloud providers, and data center operators. Startup chipmaker d-Matrix was mentioned as one of the companies that has shown interest in the initiative.

How to try it

Developers looking to try Mojo 1.0 can download the compiler, standard library, and documentation from mojolang.org. Installation is straightforward on macOS and Linux, while Windows users will need to go through WSL for now.

Signing up for Modular Cloud lets users pick a public model from the catalog and call it directly via the shared endpoint, or upload their own model for dedicated deployment. Because the service uses an OpenAI-compatible API, existing code built with the OpenAI SDK can be reused simply by swapping the endpoint address — a detail likely to matter most in practice.

To try Modular Cloud now, create an account and browse the model catalog at the Modular Cloud free sign-up page.

Editor's take

At first glance, Modular fully open-sourcing Mojo while simultaneously being absorbed into Qualcomm might seem contradictory, but the sequence makes sense. Given that the acquirer is a specific chipmaker, leaving the language and compiler closed would have risked cementing a perception of Mojo as a "Qualcomm-only tool." Opening even the compiler under Apache 2.0, so that anyone can port it to a new platform, reads as a signal that Modular intends to remain a neutral standard even after the acquisition.

This pattern isn't unprecedented. Whenever a chipmaker has acquired a software standard, that standard's survival has always come with a condition: it must keep supporting the acquirer's competitors. Modular, which already supported NVIDIA and AMD GPUs, has now added Qualcomm's own chips (Cloud AI 100 Ultra, Dragonfly) alongside Amazon Trainium and Google TPUs all at once — putting that condition into practice.

For companies in Korea, the claimed reduction in engineering effort is worth watching. HTEC's reported success in adding TPU support with a small team in just a few months could serve as a reference point for semiconductor startups or cloud operators developing their own accelerators and trying to estimate porting costs. That said, since the figure comes from Modular itself, actual porting difficulty may vary depending on the specific hardware involved.

Two things are worth watching over the coming months: how quickly Modular Cloud moves Trainium and TPU support into real production use, and which companies end up joining the alliance program. With a major model developer like MiniMax already proving performance through live production traffic, the next question is whether this platform can establish itself as a shared industry foundation rather than the tool of a single chip company.