I Built JARVIS to Work Completely Offline (INSANE Results!)

Summary

The video demonstrates a personal assistant called Jarvis that runs AI models, voice interaction, and memory locally on a computer. The creator presents this setup as a way to reduce reliance on cloud AI services, subscriptions, and API fees. Local processing is described as more private, although tools such as Gmail, Google Calendar, and Notion may still require an internet connection.

Features shown

• Switching between local and cloud models: Jarvis can be asked to switch between models, including examples from the Qwen and Gemma families, or return to a cloud model using an Anthropic API. The creator says Jarvis checks whether a requested local model is available before switching.
• Local voice: Jarvis can switch from a cloud voice service to a locally run voice clone using the open-source Voice Studio tool. Starting the voice app for the first time may add a short delay.
• Local “second brain”: Jarvis can search local files and retrieve information from folders. The demonstration shows it finding material about a community.
• Tools and integrations: The assistant can connect to services such as Gmail, Google Calendar, and Notion, and can be given additional tools. These online integrations are separate from the fully local model and voice setup.

Hardware and model guidance

The creator identifies unified memory as the main hardware consideration for running local models on Macs, because the CPU, GPU, and neural engine share it. More memory allows larger models and may improve responsiveness. Storage is less limiting because external drives can be used.

The video offers a rough guideline of about 0.6 GB of memory per billion model parameters, while noting that actual results depend on the machine and model. Examples include:

• A 32 GB Mac Mini or similar machine for smaller models; some models around 20–27 billion parameters are described as workable.
• Around 128 GB for a 70-billion-parameter model.
• Around 256 GB for a model near 125 billion parameters.

These figures are estimates, not guarantees. The video also mentions Windows systems with Nvidia GPUs and other powerful machines as alternatives. Some model and chip names in the auto-generated subtitles may be inaccurate.

Guides and setup walkthroughs

1. Ollama model installation: Find a model on Ollama, copy its command, and use commands such as ollama pull <model> to download it and ollama run <model> to test it.
2. Local voice cloning: Download Voice Studio and follow the installation instructions for the relevant device, available through its project page or GitHub.
3. Building Jarvis: The creator points viewers to a free prompt pack and guide for building their own assistant. A paid community offers a preconfigured Jarvis package, described as version 8.1, for a quicker setup.
4. Hardware selection: The video compares Mac Mini and Mac Studio options and directs viewers to a guide that also covers Windows and Nvidia-based systems.

Nature of the video

The video is a product demonstration and setup guide, not an independent review. It showcases model and voice switching, file retrieval, and the creator’s assistant interface.

Main speaker/source: The video’s creator, Zubair Trabzada, narrates the demonstration and setup guidance. Jarvis appears as the demonstrated AI assistant, not as a separate human speaker.