FrameBuff

Docs / AI & GPU

Local models and overnight renders

FrameBuff runs AI models on your own hardware. You see a review of every model before anything is downloaded, and everything a model makes stays on your machine.

Install a model

Open Models & harnesses. Paste a link to a model on Hugging Face or GitHub, or pick one from the Catalog.

  1. Fetch. FrameBuff reads the link and the file list. Nothing is downloaded yet.
  2. Review. You see what the model is, what it can do, how big it is and where it will go.
  3. Install. FrameBuff downloads only the weight files the review named. It checks each file's size, and its SHA-256 checksum when the host publishes one. Code that ships beside the weights is listed and never fetched.

The Catalog checks each model against your machine's memory, free GPU memory and processor before you pick it. Downloads larger than 40 GB are refused unless you raise Largest download.

Models already on your disk can be adopted with no download, under Settings → Earlier data → Models already on this machine. A file whose checksum matches the catalog is marked verified. Any other file is adopted only when you confirm it.

Serve and stop

Serving a chat model starts it as a job you can watch and stop. Models larger than Largest model for the CPU (6 GB by default) go onto your graphics card when one answers. Smaller models stay on the processor.

Chat models run on llama.cpp and need the GGUF format. You install llama.cpp yourself from its releases, in Settings → Paths → Runtimes.

Roles

FrameBuff has four roles: Chat, Small tasks, Coding and Vision. Each role points at one model, and the rest of FrameBuff asks for a role. Each role also unloads its model after it has been idle for a while: Chat after 15 minutes, Coding after 30, Vision after 2, and Small tasks never. An unload waits for running work and never stops a model mid-answer.

A role can also use a model on another machine that is lent to you. See Lend your GPU.

Images, video and music

The Catalog also lists the engines that run other kinds of models, such as ComfyUI for images and video and MusicGen for music. Each engine installs into a folder of its own after you confirm it. Engines load their models offline, and every result stays on your machine.

Render while you sleep

The Make a video app turns a sentence into a short clip. Choose a length of 2 to 10 seconds, then choose Now or Tonight. Tonight asks you once on the approval bar, then renders in the night window, 01:00 to 05:00 by default, while nobody needs the card. It renders through ComfyUI on your own graphics card, and the clip stays on your machine.

Online models, if you want them

Online models are off by default. If you turn on Online mode, you can pick models from OpenRouter from a short list you choose. A monthly spending cap applies, 5 dollars until you set one, and a paid call over the cap is refused. Your OpenRouter key is kept in the vault.

You can also connect a model server a friend runs, such as Ollama or llama.cpp. FrameBuff shows a warning first, because outside FrameBuff nothing is encrypted end to end. A key is required, and it is only ever sent to the address you typed.