Megastructure Arc · July 12, 2026

Arc 1.3.0

Arc 1.3.0 accelerates bundled Local AI on Windows with Vulkan, fixes Mac launch, and makes first-run settings safer.

Previous release

Faster Local AI, fixed on Mac

Windows 10/11 (64-bit), or macOS 11+ on Apple Silicon.

Highlights

  • Windows now ships the official llama.cpp Vulkan runtime. Supported NVIDIA, AMD, and Intel GPUs can accelerate models, with the same CPU fallback when no working GPU is available.
  • Apple Silicon Local AI launches correctly again: Arc now includes every dylib required by the upstream llama-server, and the build fails early if the staged runtime cannot start.
  • A new “Use bundled runtime” control restores Arc’s shipped llama-server path after a custom override.

Fixes and refinements

  • New Local AI profiles default to a 32,768-token context window so first launch fits more reliably on 8–12 GB GPUs. Existing profiles keep their chosen setting.

System requirements

Windows 10/11 (64-bit), or macOS 11+ on Apple Silicon.