August 2026 Summaries

Modular announced that its heterogeneous AI computing platform is production-ready, reportedly serving billions of tokens per minute in enterprise deployments, and introduced several initiatives to broaden its availability and hardware support. Mojo 1.0, including its compiler and tooling, is now fully open source under the Apache 2.0 license, with native Windows support being developed in collaboration with Microsoft. Modular Cloud is generally available with OpenAI-compatible shared endpoints and dedicated deployments, and the company identifies MiniMax as a flagship customer using it to serve the M3 model at scale. The platform now supports NVIDIA and AMD GPUs as well as AWS Trainium, Google TPUs, Qualcomm Cloud AI 100 Ultra, and Qualcomm Dragonfly accelerators, allowing models to use common APIs and abstractions across hardware types. Modular also plans to make MAX source-available, remove device-use restrictions from its license, and establish an industry alliance program intended to let hardware vendors, cloud providers, model developers, and other partners extend and optimize the platform.

Mojo 1.0 has released its complete compiler, toolchain, and related source code under the permissive Apache 2.0 license with LLVM exceptions, making the general-purpose language fully open source after four years of development with a previously closed compiler. Designed to apply modern compiler and programming-language research to GPUs, AI accelerators, and other advanced hardware, Mojo had already opened its standard library, kernel code, and development discussions before this broader release. The source is available through Modular’s GitHub repository and can be built locally with Bazel, allowing developers to compile Mojo, modify its standard library, and run tests, while a prebuilt nightly compiler option is available for faster use cases. Although the standard library has accepted outside contributions since 2024, contributions to the compiler and tooling are not yet open, with Modular aiming to enable them by the end of the year.

Modular announced that Qualcomm Technologies’ data center AI accelerators, beginning with the Cloud AI 100 and extending to the Dragonfly AI 200 and future chips, are being integrated into its MAX and Mojo software platform, allowing developers to target Qualcomm NPUs with the same model code and tooling used for NVIDIA and AMD GPUs. The collaboration brought the Cloud AI 100 Ultra online through a native LLVM Hexagon-based backend rather than an adapter layer, accommodating its SIMD-oriented execution model, explicit DMA-managed memory, vector and matrix engines, and multi-chip architecture. After initially enabling optimized kernels and an end-to-end GPT-2 pipeline, the teams reportedly progressed to serving the Gemma 4 31B model across four devices in under six months, with support for tensor parallelism, paged attention, speculative decoding, and a mixture-of-experts variant. Modular says the work demonstrates its portability-by-design approach, in which shared compiler, runtime, graph, model, and serving components reduce the effort required to support new hardware, while ongoing work aims to optimize communications, extend model support, and complete integration for the newer Dragonfly AI 200 platform.

Mojo 1.0 marks the language’s transition from rapid early development to a stable, production-ready foundation intended to support long-term projects and ecosystem growth. Modular says the language, first released in 2023, is already used internally for its MAX and Modular Cloud infrastructure, while community contributors have submitted more than 1,100 pull requests to the open-source standard library. Release 26.5 finalizes language simplifications such as consistent variable declarations, unified closures, a single Pointer type, and clearer terminology, while adding Python-style lambdas, a more reliable language server, improved AI skills, memory-reference invalidation diagnostics, and more actionable “where” clause errors. Future plans include asynchronous programming, pattern matching, unions, continued developer-experience improvements, additional open-sourcing of Mojo and MAX components, and the planned release of the Mojo compiler and toolchain as open source in 2026. The accompanying MAX 26.5 update introduces modular installation options, support for GLM-5.2 and Nemotron-H models, Kimi 2.5 compatibility with Module V3, and open-source agent skills for model lifecycle work.