Mojo 1.0 ships: one stable language for CPU, GPU and NPU code
After three years of churn, Modular locks the API — GPU kernels hit ~87% of CUDA on an H100, and Tencent drops an open one-prompt 3D-world builder.
Copy markdown
Three years of churn, over
Mojo 1.0 (Modular 26.5) freezes a production-ready API — changes stay additive through the 1.x line, C++-style, so code you write today keeps compiling. Grab it with `uv pip install --upgrade mojo`; nearly 200 contributors have landed 1,100+ PRs on the open standard library.
One source, CPU to GPU to NPU
Built on MLIR, a single Mojo file targets CPU, GPU, TPU and ASIC. Oak Ridge benchmarked its GPU kernels at ~87% of CUDA on an NVIDIA H100 for memory-bound work — portable kernels without dropping into vendor-locked C++.
It's a Qualcomm language now
Qualcomm closed its roughly $3B Modular buy on Jul 28, so Hexagon NPUs and Dragonwing silicon become addressable from the same Mojo codebase. The compiler is slated to go open source by year-end, possibly at ModCon on Aug 18.
Elsewhere: one prompt, a whole 3D world
Tencent's Hunyuan team unveiled WorldClaw, an agentic framework that turns a single prompt into an explorable, editable open-world 3D scene with region-aware terrain — weekend fuel for game and sim builders, with a paper and project repo out now.