The Mojo language (by Modular, now Qualcomm) is now open-source

Four and a half years ago, Modular made a bet: AI would not tally connected 1 benignant of silicon forever, and the package stack would request to beryllium rearchitected for a world of heterogeneous hardware and progressively analyzable AI workloads.

At the ModCon keynote this morning, we showed what that stake has become. The Modular Platform is now production-ready, serving billions of tokens per infinitesimal and powering existent endeavor deployments. We’re opening much of the level to the ecosystem, extending it crossed wholly caller classes of hardware, and bringing awesome manufacture partners on pinch us.

More specifically, Modular and Qualcomm announced:

• Mojo 1.0 is now afloat unfastened source nether an Apache 2.0 license.
• Modular Cloud is publically available, serving flagship customers for illustration MiniMax.
• Modular Platform now supports AWS Trainium, Google TPUs, and the Qualcomm Cloud AI 100 and Qualcomm Dragonfly accelerators alongside CPUs and GPUs. Learn much astir the Qualcomm Dragonfly bringup activity successful our blog post.
• Native Windows support is coming to Mojo, acknowledgment to a collaboration pinch the Microsoft Windows team.
• The MAX licence nary longer contains instrumentality usage restrictions, and MAX will beryllium source-available pinch an unfastened confederation program, so the broader ecosystem tin build the level pinch us.

Heterogeneous compute is here, and it has a single, unfastened package platform.

Mojo 🔥 is now unfastened source

Last week, we announced that Mojo reached 1.0, providing developers a stable, production-ready instauration they tin build connected for the agelong term. Alongside a scope of caller features, the astir important portion of 1.0 is the stableness guarantee: the codification you constitute coming won’t break retired from nether you.

Today we're going a measurement further. The full Mojo connection is now unfastened root nether the unrestricted Apache 2.0 license, which intends the compiler and each tooling are afloat unfastened source. You tin widen the language, bring it to caller platforms, and build immoderate applications you want connected apical of it. This continues the progressive opening of our stack that began pinch the Mojo modular room successful 2024 and the MAX kernels successful 2025, and it will proceed from here. To study much astir Mojo and contribute, sojourn mojolang.org.

Windows support for Mojo

Mojo has supported macOS and Linux for years, and Windows developers person been capable to usage it done WSL. Native Windows support has ever been 1 of our astir communal requests from developers.

Millions of developers build connected Windows each day, crossed an tremendous scope of applications and workloads. We judge Mojo tin person a meaningful effect crossed that ecosystem. Bringing Mojo to Windows the correct measurement requires heavy expertise successful the platform, which is why we’re delighted that the Microsoft Windows squad sees the aforesaid opportunity we do — and that we’re moving together to make it happen.

"We're excited to spot Mojo coming to Windows and the opportunities it creates for developers moving crossed systems and AI. Millions of developers build connected Windows each day, and we're committed to helping them entree the devices and technologies they request connected the level they choose."

– Logan Iyer, CVP, Windows Platform + Developer

Introducing Modular Cloud

Modular Cloud is wherever the afloat Modular stack comes together arsenic a accumulation service. It gives developers nonstop entree to Modular’s industry-leading conclusion capacity while abstracting distant the complexity of deploying, optimizing, and operating models crossed heterogeneous infrastructure.

Modular Cloud is mostly disposable astatine console.modular.com, serving celebrated unfastened root models connected the Modular stack done shared endpoints and dedicated deployments. Shared endpoints are OpenAI-compatible pinch pay-per-token pricing, while dedicated deployments tally connected our compute aliases your own, connected reserved isolated instances.

Modular Cloud has been softly serving OpenRouter postulation for the past fewer months nether the sanction ModelRun, wherever its endpoints person consistently classed astatine aliases adjacent the apical of the level for latency and throughput connected accumulation traffic. Artificial Analysis, an independent benchmarking firm, tells the aforesaid story.

MiniMax

MiniMax is simply a flagship endeavor customer of Modular Cloud, moving M3 connected a dedicated Modular deployment that serves its accumulation postulation astatine billions of tokens per minute.

Serving M3 efficiently astatine standard presents a unsocial systems challenge. It combines a 1M-token discourse window, autochthonal multimodality, and MiniMax Sparse Attention (MSA) — a caller sparse-attention architecture that selectively attends to applicable KV blocks, reducing the compute required arsenic discourse scales.

Delivering state-of-the-art capacity required optimization crossed the stack: implementing M3 natively successful MAX, building and tuning specialized MSA kernels, and optimizing the deployment astir MiniMax’s real-world postulation patterns.

Beyond GPUs: Trainium, TPUs, and Dragonfly

Modular Cloud is already serving accumulation workloads connected NVIDIA and AMD GPUs. Today, we’re expanding that hardware support beyond GPUs and into civilization AI accelerators — a overmuch much demanding trial of the platform’s portability.

We’ve added support for AWS Trainium, Google TPUs, Qualcomm Cloud AI 100 Ultra and Qualcomm Dragonfly. Each runs done the aforesaid Modular Platform, pinch the aforesaid modeling APIs, serving workflows, programming language, and halfway abstractions. Developers tin writer a exemplary erstwhile and bring it to wholly different hardware architectures without rebuilding the package stack astir it.

Just arsenic importantly, we brought up each of these platforms pinch a fraction of the engineering effort traditionally required to alteration caller AI hardware — much than 10x simplification successful engineering effort.

Over the coming months, we’ll beryllium bringing these caller hardware platforms into accumulation and making them disposable done Modular Cloud.

MAX: A communal foundation, built pinch the ecosystem

AI hardware invention is accelerating, but awesome hardware only matters if developers tin usage it. Too overmuch of that invention is still gated down package stacks written for 1 architecture. The manufacture needs a communal package instauration instead: constitute a exemplary erstwhile and scope each accelerator, take hardware based connected capacity and economics alternatively than which stack happens to support it, and fto vendors compete connected the merits of their silicon. That is what we are building Modular Platform to be.

Joining Qualcomm reinforces that goal. Modular Platform will proceed supporting and optimizing for a wide scope of hardware, including hardware that competes straight pinch Qualcomm Technologies’ platforms. The opportunity successful beforehand of the ecosystem is overmuch bigger than immoderate azygous vendor's roadmap, and a instauration only useful if everyone tin guidelines connected it.

To build that foundation, we’re taking connected 2 important initiatives:

We’re opening up MAX. We’re evolving MAX’s licensing exemplary and expanding root entree truthful developers, enterprises, hardware vendors, and partners tin build connected the platform, widen it, and lend back.

We’re moving connected building an confederation programme for the ecosystem. We’re moving toward an manufacture confederation programme spanning hardware vendors, exemplary providers, unreality companies, and data-center operators. The extremity is to springiness partners a nonstop domiciled successful integrating MAX, optimizing it for their platforms, and helping style wherever the Modular Platform goes next.

HTEC has already shown what that looks for illustration successful practice. Their engineers brought up Google TPU support connected Modular Platform themselves successful only a fewer months pinch only a fewer engineers, pinch america successful a supporting domiciled alternatively than driving the integration. This is clear validation of what we’re building: a instauration the ecosystem tin widen independently.

And different startup hardware vendors spot the aforesaid request successful the market. d-Matrix is 1 of them:

"d-Matrix's and Modular's shared committedness to heterogeneous computing is underpinned by communal support of unfastened standards. As we participate the era of disaggregated heterogenous compute, unfastened standards tin accelerate deployment of GPUs, CPUs, and XPUs moving together to thrust efficiencies astatine scale, and it's why our business is specified a earthy fit."Sid Sheth, Founder & CEO of d-Matrix

Try it today

Everything you request to get started pinch unfastened root Mojo 1.0 is disposable now astatine mojolang.org. Modular Cloud is unrecorded astatine console.modular.com. And if you're pinch america coming successful San Francisco for ModCon, this afternoon's tech talks spell deeper connected everything above.

AI is moving excessively quickly for each institution to support rebuilding the aforesaid infrastructure underneath it. Models go larger, serving becomes much distributed, and the hardware becomes much heterogeneous. Open horizontal platforms person reset industries before, but only erstwhile the ecosystem showed up to build them together. So if you activity connected hardware, models, infrastructure, aliases applications, travel build this instauration pinch america – scope retired to america astatine [email protected].