How local AI works on UtilityHub

The short answer

Writing AI runs in a WebGPU browser after an explicit model download. Image generation, transcription and advanced media tools let you choose the Windows companion or a configured secure cloud job; the screen identifies the processing location before upload.

Steps to follow

1. Check whether the tool uses browser AI or the companion.
2. Meet the stated hardware and setup requirements.
3. Allow model downloads explicitly, then submit your inputs.
4. Review the output and clear or export results as needed.

Browser writing requirements

The writing tools use Qwen 1.5B through WebLLM, with about a 1 GB download and about 2 GB of GPU memory. The first load can take several minutes. Inputs are limited to 1,500 characters and responses to 320 model tokens. A length-limited response reports an error, so request shorter sections when needed.

The desktop companion is a separate service

Private mode uses the companion at 127.0.0.1:8765 with an isolated Python environment, FFmpeg and LibreOffice; image generation also requires a supported NVIDIA GPU. Cloud mode is a separate authenticated processor configured by the operator. Files are limited to 50 MB each and 100 MB total, with additional operation-specific limits.

Downloads are different from inference

Model publishers receive normal requests when weights are downloaded. Inputs are not included in those downloads. The local companion deletes temporary files after its response; cloud inputs are deleted after processing and results expire automatically. Browser model caches remain until removed. URL-checking tools contact the explicit public URL you enter.

Review every generated result

Local execution does not make a model accurate. Writing can invent claims, captions can misdescribe images and transcripts can miss words. Beta labels remain visible. Review output before publishing or relying on it, and supply specific source facts rather than asking the model to guess.