Local Cognitive 0.1.1

The first public release of Local Cognitive: private chats, agents and workflows with models on your own computer, or on your own GPU server through Remote.

Desktop

• macOS (Apple silicon). Signed and notarized. New versions install from the app: the update button in the title bar, or Settings → About.
• Windows (x64), new.
  • Models run on any GPU through Vulkan. On a PC with an NVIDIA GPU, the app offers the CUDA build: about 515 MB, after NVIDIA's license.
  • The installer is not signed yet. Windows asks once (More info → Run anyway), and the app announces new versions with a link to this page.
• Clearer model errors. A model file that the bundled runtime cannot read now says why, for example a format made for a modified llama.cpp. The error can be dismissed.
• Split models. Models made of several GGUF files install and import.
• GPU memory. Models shows the GPU's memory beside RAM.
• Server models in answers. Answers from a server name the server's model.
• Menus. A click anywhere closes an open menu, the title bar included.

Server (Linux)

• Several NVIDIA GPUs. The service sees every NVIDIA GPU: each one gets its device file when the service starts.

• Install:

  curl -fsSL https://github.com/Ilyaberdar/local-cognitive-AI-system/releases/latest/download/install.sh | sudo bash
  

• Update:

  sudo local-cognitive-server update
  

llama.cpp b10809 CUDA 12.8.1 overlay (linux-x64)

llama.cpp b10809 CUDA 12.8.1 overlay for linux-x64.

Adds libggml-cuda.so (architectures 50-virtual;61-real;70-virtual;75-real;80-real;86-real;89-real;90-real;90-virtual;120a-real) and the NVIDIA CUDA runtime and cuBLAS libraries to the upstream linux-x64 build.
Requires an NVIDIA driver 570 or newer. NVIDIA components are redistributed under the CUDA EULA; see CUDA_NOTICE.txt in the archive.

Paste manifest-fragment.json into resources/llama/runtime-manifest.json, then run npm run prepare:llama:server.