Modular

Modular builds AI infrastructure around Mojo, a Python-oriented systems programming language, and MAX, a model compiler, runtime and serving framework. Its central engineering decision is to connect custom computation with the execution engine: operations written in Mojo can participate in compiler analysis and graph fusion rather than remain separate extensions. Founded by Chris Lattner and Tim Davis in early 2022, Modular became part of Qualcomm in July 2026; Mojo, MAX and Modular Cloud continue as products and brands.[2][4][9][8]

The integrated stack grew out of the founders' earlier work together at Google. Their initial products paired inference infrastructure with a language for expressing the computation inside it.[1][2]

Lattner and Davis met at Google and founded Modular in early 2022. In his September 2023 interview, Lattner recalled dividing the work: he built the engineering team and technical foundation, while Davis developed the product and business sides and interviewed roughly 50–100 companies about their infrastructure problems. Their diagnosis was that hardware-specific software, training frameworks and layers of deployment tooling had left AI difficult to put into production.[1][2][3]

Mojo emerged from a practical difficulty within the compiler work. Lattner recalled, “we were writing this very low level MLIR stuff by hand.” The team needed usable syntax and chose a full programming language over an embedded domain-specific language. In his account, debugger and tooling limitations helped determine that choice; Python's community and programmers' existing habits guided the language's design. Mojo and the engine therefore developed together, with the language providing a way to express the engine's kernels.[3][2]

The May 2, 2023 launch, authored by Lattner, Davis and Eric Johnson, introduced the Modular Inference Engine and Mojo. The engine offered Python and C/C++ APIs for deploying PyTorch and TensorFlow models, including dynamic shapes and custom operators. Mojo used MLIR and combined Python-oriented syntax and access to Python libraries with systems programming features, static types and compile-time metaprogramming. Initial access was limited: the engine went to early-access partners, while Mojo was available through a JupyterHub-based playground.[2]

At ModCon on December 4, 2023, Modular introduced MAX—Modular Accelerated Xecution—combining MAX Engine, formerly AI Engine, with MAX Serving and Mojo. Developer and enterprise editions were announced for early 2024. Modular also announced an AWS partnership to bring MAX Enterprise Edition to AWS production services and a technology partnership with NVIDIA for accelerated CPU and GPU computing.[4]

Compiler-visible custom operations are a defining mechanism of MAX. Developers can write operations in Mojo for native analysis and graph fusion, without building and loading separate C++ or CUDA shared libraries. The Graph API supports model graphs, custom preprocessing and postprocessing, graph optimization and kernel generation. This gives the compiler access to specialized computation alongside the surrounding model operations.[4]

The MAX 24.6 release on December 17, 2024 introduced a MAX GPU preview focused on generative AI. MAX Engine supplied the model compiler, runtime and Mojo kernels for NVIDIA GPUs; MAX Serve supplied Python-native request batching and scheduling for LLM applications. An OpenAI-compatible client API and Docker deployment connected the stack to familiar application and cloud workflows. Its independence from CUDA kernels did not imply independence from NVIDIA drivers or every CUDA-related dependency.[5]

Hardware coverage expanded with MAX v25.4 on June 18, 2025. It supported model acceleration on AMD MI300X and MI325X GPUs, while initial RDNA3 support covered GPU programming with some required model kernels still missing. The release added max.torch.CustomOpLibrary, illustrated with a grayscale operation, and previewed importing Mojo code from Python. The PyTorch custom-operations tutorial remained beta. It also removed TorchScript and torch MLIR model support, changing the framework coverage advertised in earlier generations.[7]

In his June 2025 interview, Lattner described a transition from investigating how to build an alternative heterogeneous-compute stack to managing engineering work once it functioned end to end: “It's not R&D anymore. It's an engineering problem.” He described six-week releases and work on hardware support, function calling and token sampling, while acknowledging that the system remained early and had bugs.[6]

Lattner also recounted a hackathon in which a four-person team, new to Mojo and GPU programming, used the inference infrastructure to build an Adam optimizer, training kernels and a simple backpropagation system in one day. He credited AI coding tools as part of the result. The experiment demonstrated, in his account, how participants could extend infrastructure built for inference into training computation; it was an early demonstration rather than a production training platform.[6]

By September 2025, Modular described a broader platform spanning kernels through distributed serving. Alongside Mojo and MAX, Mammoth provided a Kubernetes-native control plane and router for large-scale AI serving, with multi-model management, prefill-aware routing and disaggregated compute and cache. The company's financing announcement identified cloud scaling and expanded cloud and edge hardware coverage as uses for its new capital.[17]

Modular's first financing round raised $30 million in 2022. On August 24, 2023, it announced a further $100 million led by General Catalyst, with existing investors GV, SV Angel, Greylock and Factory participating. On September 24, 2025, it announced $250 million in its third round, led by Thomas Tull's US Innovative Technology fund, with DFJ Growth joining and existing investors participating. Modular reported total capital raised of $380 million and a $1.6 billion valuation.[16][17]

In September 2025, Modular reported more than 130 employees, a San Francisco Bay Area headquarters and a footprint across North America, the United Kingdom and Europe. Earlier, at ModCon in December 2023, it reported a community of 150,000 developers and more than 56,000 Mojo SDK downloads since the SDK's September availability. These were company-reported staffing and developer-community measures, rather than counts of paying customers.[17][4]

Qualcomm completed its acquisition of Modular on July 28, 2026, and announced completion the following day. Its quarterly filing valued the transaction at approximately $3.1 billion using Qualcomm's acquisition-date stock price, with consideration primarily comprising 18 million common shares. That valuation included service-dependent executive shares and differed from a finalized accounting purchase price.[9][8]ⁿ¹

Mojo, MAX and Modular Cloud continued as products and brands. Following the acquisition, Modular's company page identified Lattner as Qualcomm's EVP of AI Software and Platforms and Davis as SVP and general manager of Modular. Modular subsequently committed to continuing support for hardware that competes with Qualcomm's platforms.[8][1][10]

Mojo 1.0 arrived on August 11, 2026. Modular explained that frequent language changes had made community projects difficult to maintain, prompting an emphasis on stability. Most core language features were stable, while the standard library began with a deliberately small set of individually stable APIs. The release itself included breaking changes, generally accompanied by deprecated aliases and compiler fix-its.[14][15]

The stability policy applies to source compatibility, while the application binary interface remains unstable. A stable struct signature does not automatically stabilize its members, and unstable features—including async/await—can change. Critical-issue exceptions remain possible; even permitted additions of defaulted type parameters can break code that explicitly unbinds every parameter using ellipsis.[13]ⁿ²

At ModCon on August 18, 2026, Modular opened Mojo's compiler and tooling under Apache 2.0 with LLVM Exceptions, extending the opening of its standard library in 2024 and MAX kernels in 2025. The repository contains the Mojo compiler and standard library, MAX accelerator library, Python model pipelines, examples and OpenAI-compatible inference server. Contributions are accepted for several components, but not yet for the Mojo compiler. Modular also announced forthcoming native Windows support in collaboration with Microsoft's Windows team.[10][12][11]

ModCon 2026 also marked Modular Cloud's general availability. Shared endpoints offer OpenAI-compatible access with pay-per-token pricing; dedicated deployments use reserved instances on Modular or customer compute. Modular said the service had carried OpenRouter traffic under the name ModelRun for several months and named MiniMax as a flagship enterprise customer.[10][12]

For MiniMax's M3 deployment, Modular described implementing the model natively in MAX, tuning specialized MiniMax Sparse Attention kernels, and optimizing serving around production traffic. M3 combines a one-million-token context window, multimodality and sparse attention that selects relevant KV blocks. Modular reported that the dedicated deployment served billions of tokens per minute.[10]

The August announcements added AWS Trainium, Google TPUs, Qualcomm Cloud AI 100 and Dragonfly to the platform's hardware coverage, with different levels of maturity. The accompanying FAQ specified that AI100 support extended through Modular Cloud, while TPU and Trainium could execute Gemma models through MAX but lacked Cloud integration. Further production and Cloud expansion remained planned for the coming months. Modular credited HTEC engineers with bringing up Google TPU support, with Modular in a supporting role.[10][12]

Modular announced broader MAX source availability and work toward an industry alliance for hardware vendors, model providers, cloud companies and data-center operators. The revised Community License, dated August 18, removed earlier device-use restrictions while retaining other conditions. Separately Apache-licensed components keep their own terms; MAX usage and distribution remain subject to the Community License.[10][12][11]

The license distinguishes redistribution of object-code derivatives from redistributed Applications, which must add material functionality. It permits AI tools to analyze, improve and explain MAX or develop interoperating software, but restricts using MAX as input to produce a substitute implementation. Legal-entity providers selling hosted or managed AI services to other legal entities must meet branding requirements, including prior written approval for marked materials.[12]ⁿ³