Four and a half years ago, Modular made a bet: AI would not run on one kind of silicon forever, and the software stack would need to be rearchitected for a world of heterogeneous hardware and increasingly complex AI workloads.
At the ModCon keynote this morning, we showed what that bet has become. The Modular Platform is now production-ready, serving billions of tokens per minute and powering real enterprise deployments. We’re opening more of the platform to the ecosystem, extending it across entirely new classes of hardware, and bringing major industry partners along with us.
More specifically, Modular and Qualcomm announced:
Heterogeneous compute is here, and it has a single, open software platform.
Last week, we announced that Mojo reached 1.0, providing developers a stable, production-ready foundation they can build on for the long term. Alongside a range of new features, the most important part of 1.0 is the stability guarantee: the code you write today won’t break out from under you.
Today we're going a step further. The entire Mojo language is now open source under the unrestricted Apache 2.0 license, which means the compiler and all tooling are fully open source. You can extend the language, bring it to new platforms, and build whatever applications you want on top of it. This continues the progressive opening of our stack that began with the Mojo standard library in 2024 and the MAX kernels in 2025, and it will continue from here. To learn more about Mojo and contribute, visit mojolang.org.
Mojo has supported macOS and Linux for years, and Windows developers have been able to use it through WSL. Native Windows support has always been one of our most common requests from developers.
Millions of developers build on Windows every day, across an enormous range of applications and workloads. We believe Mojo can have a meaningful impact across that ecosystem. Bringing Mojo to Windows the right way requires deep expertise in the platform, which is why we’re delighted that the Microsoft Windows team sees the same opportunity we do — and that we’re working together to make it happen.
"We're excited to see Mojo coming to Windows and the opportunities it creates for developers working across systems and AI. Millions of developers build on Windows every day, and we're committed to helping them access the tools and technologies they need on the platform they choose."
– Logan Iyer, CVP, Windows Platform + Developer
Modular Cloud is where the full Modular stack comes together as a production service. It gives developers direct access to Modular’s industry-leading inference performance while abstracting away the complexity of deploying, optimizing, and operating models across heterogeneous infrastructure.
Modular Cloud is generally available at console.modular.com, serving popular open source models on the Modular stack through shared endpoints and dedicated deployments. Shared endpoints are OpenAI-compatible with pay-per-token pricing, while dedicated deployments run on our compute or your own, on reserved isolated instances.
Modular Cloud has been quietly serving OpenRouter traffic for the past few months under the name ModelRun, where its endpoints have consistently ranked at or near the top of the platform for latency and throughput on production traffic. Artificial Analysis, an independent benchmarking firm, tells the same story.
MiniMax is a flagship enterprise customer of Modular Cloud, running M3 on a dedicated Modular deployment that serves its production traffic at billions of tokens per minute.
Serving M3 efficiently at scale presents a unique systems challenge. It combines a 1M-token context window, native multimodality, and MiniMax Sparse Attention (MSA) — a novel sparse-attention architecture that selectively attends to relevant KV blocks, reducing the compute required as context scales.
Delivering state-of-the-art performance required optimization across the stack: implementing M3 natively in MAX, building and tuning specialized MSA kernels, and optimizing the deployment around MiniMax’s real-world traffic patterns.
Modular Cloud is already serving production workloads on NVIDIA and AMD GPUs. Today, we’re expanding that hardware support beyond GPUs and into custom AI accelerators — a much more demanding test of the platform’s portability.
We’ve added support for AWS Trainium, Google TPUs, Qualcomm Cloud AI 100 Ultra and Qualcomm Dragonfly. Each runs through the same Modular Platform, with the same modeling APIs, serving workflows, programming language, and core abstractions. Developers can author a model once and bring it to entirely different hardware architectures without rebuilding the software stack around it.
Just as importantly, we brought up each of these platforms with a fraction of the engineering effort traditionally required to enable new AI hardware — more than 10x reduction in engineering effort.
Over the coming months, we’ll be bringing these new hardware platforms into production and making them available through Modular Cloud.
AI hardware innovation is accelerating, but great hardware only matters if developers can use it. Too much of that innovation is still gated behind software stacks written for one architecture. The industry needs a common software foundation instead: write a model once and reach every accelerator, choose hardware based on performance and economics rather than which stack happens to support it, and let vendors compete on the merits of their silicon. That is what we are building Modular Platform to be.
Joining Qualcomm reinforces that goal. Modular Platform will continue supporting and optimizing for a broad range of hardware, including hardware that competes directly with Qualcomm Technologies’ platforms. The opportunity in front of the ecosystem is much bigger than any single vendor's roadmap, and a foundation only works if everyone can stand on it.
To build that foundation, we’re taking on two important initiatives:
We’re opening up MAX. We’re evolving MAX’s licensing model and expanding source access so developers, enterprises, hardware vendors, and partners can build on the platform, extend it, and contribute back.
We’re working on building an alliance program for the ecosystem. We’re working toward an industry alliance program spanning hardware vendors, model providers, cloud companies, and data-center operators. The goal is to give partners a direct role in integrating MAX, optimizing it for their platforms, and helping shape where the Modular Platform goes next.
HTEC has already shown what that looks like in practice. Their engineers brought up Google TPU support on Modular Platform themselves in only a few months with only a few engineers, with us in a supporting role rather than driving the integration. This is clear validation of what we’re building: a foundation the ecosystem can extend independently.
And other startup hardware vendors see the same need in the market. d-Matrix is one of them:
"d-Matrix's and Modular's shared commitment to heterogeneous computing is underpinned by mutual support of open standards. As we enter the era of disaggregated heterogenous compute, open standards can accelerate deployment of GPUs, CPUs, and XPUs working together to drive efficiencies at scale, and it's why our partnership is such a natural fit."
Sid Sheth, Founder & CEO of d-Matrix
Everything you need to get started with open source Mojo 1.0 is available now at mojolang.org. Modular Cloud is live at console.modular.com. And if you're with us today in San Francisco for ModCon, this afternoon's tech talks go deeper on everything above.
AI is moving too quickly for every company to keep rebuilding the same infrastructure underneath it. Models become larger, serving becomes more distributed, and the hardware becomes more heterogeneous. Open horizontal platforms have reset industries before, but only when the ecosystem showed up to build them together. So if you work on hardware, models, infrastructure, or applications, come build this foundation with us – reach out to us at alliance@modular.com.