Mojo 1.0 ships: AI systems language locks a stable API for production
Modular's Python-meets-Rust language is production-ready and Qualcomm-owned; plus Tencent's text-to-3D-world WorldClaw and llama.cpp's new local TTS.
Copy markdown
A stable API, at last
Mojo 1.0 freezes the core language — unified `var`, one `Pointer` type, Python-style lambdas — and pledges that 1.x changes stay additive, so production codebases stop breaking on every update. Independent ORNL tests already clocked it near 87% of CUDA's H100 throughput on memory-bound kernels, with the same code running on NVIDIA and AMD GPUs.
Now a Qualcomm company
Modular closed its ~$3.1B sale to Qualcomm on July 28, so the language you'd build on is now steered by a chipmaker. The standard library stays Apache 2.0, but the compiler is still proprietary, with an open-source pledge before the end of 2026.
Build a 3D world from a sentence
Tencent's Hunyuan3D-WorldClaw turns a text prompt into an explorable, editable open-world 3D scene, assembling terrain and objects through an agentic pipeline. The code is on GitHub and the paper is public — a weekend hook for anyone building games, sims, or spatial demos.
llama.cpp learns to talk
This week's llama.cpp adds a streaming POST /tts endpoint (text-to-speech with speaker references), support for BailingMoE 3 / Ling 3.0, and NVFP4 quantization. Local inference got faster too: ~12% higher Mamba-2 prefill on Metal, plus Vulkan gains for DeepSeek-V4.