What is local AI?

What is local AI?

Local AI is artificial intelligence that runs on your own hardware instead of on a remote server. The language model weights sit on your disk, inference executes on your CPU or GPU, and the full conversation stays in local memory.

Cloud AI (ChatGPT, Claude.ai, Gemini) sends each prompt you type to a company's servers. A language model running in their datacenter generates a reply, which travels back over the internet. Local AI removes that round trip. The model file lives on your machine, and when you ask it a question your CPU or GPU does the thinking. "Runs on your hardware" means exactly that: no cloud tenant, no API call, no provider log of what you asked.

How does local AI work?

You install a local runtime (Ollama or LM Studio), download an open-source model file once, and from then on every prompt runs inference locally. The model never phones home and no request leaves your device.

The steps in practice: the runtime downloads the model weights (typically a few gigabytes) from a public registry, caches them on disk, and exposes a local HTTP endpoint on your own machine. Applications like InnerZero call that endpoint the same way a web app would call a cloud API, except the call never touches the internet. Inference uses your GPU first (fast) and falls back to CPU (slower but workable) if the model is larger than your VRAM. InnerZero is the front end: it handles the chat window, memory, voice, knowledge packs, and tools. The heavy lifting happens in Ollama or LM Studio, which we treat as swappable engines.

What is the difference between local AI and cloud AI?

Local AI wins on privacy, cost, and offline use. Cloud AI wins on raw capability for frontier reasoning tasks. Hybrid means local by default with cloud available when a task needs it; it is what InnerZero supports via optional API keys and managed cloud.

The honest trade-off looks like this:

For a deeper walkthrough of the trade-offs, see local AI vs cloud AI.

What hardware do I need to run local AI?

Any modern laptop can run a small model (4B parameters) on CPU alone. A consumer GPU with 8 to 16 GB of VRAM comfortably runs the mid-range models most users reach for. Frontier models in the 30B to 120B range need a high-end GPU or multi-card setup.

InnerZero ships with a tier ladder so the right model gets picked automatically for your machine. Entry tier (8 to 16 GB VRAM) runs Qwen 3 8B for chat and code. Standard tier (16 to 24 GB VRAM) steps up to Qwen 3 30B for stronger reasoning. Pro tier (48 to 80 GB VRAM or a pair of cards) runs gpt-oss 120B for heavy tasks. Enterprise and frontier tiers push into multi-GPU workstations and the 235B class. For the full matrix, see the supported models page, and for a practical buying guide read hardware for local AI.

Is local AI free?

Yes. The open-source models and the runtimes that serve them are free to download and use. InnerZero itself is free forever for personal use. Your only marginal cost is the electricity each inference uses.

The realistic catch is not money, it is time and hardware. A model big enough to feel useful wants a decent GPU, which you either already have or buy once. First-time setup (installing the runtime, downloading a model) takes half an hour end to end. After that, every prompt is free. Compare this to a cloud subscription at £15 to £20 a month: local breaks even in well under a year of regular use. See AI without a subscription for the full maths.

Is local AI private?

Yes. By default, no prompt, file, or memory leaves your machine. The model is a static file; it has no network access of its own. Only you see what you type and what the AI says back.

The caveat is that "private" depends on the app wrapper you use. InnerZero is designed so local mode truly is local: no telemetry, no analytics, no crash reports, no cloud sync. Optional cloud mode exists (for tasks where you want a frontier model) but is strictly opt-in, and even then only the current prompt plus a short conversation window is sent, never your full memory or files. For the formal statement of what is and is not collected, see the privacy policy.

Can local AI work offline?

Yes. Once the model is downloaded, local AI runs with no internet connection at all. Planes, trains, remote cabins, air-gapped work laptops: all fine.

The one caveat is first-time setup. Installing the runtime and downloading the model weights does require internet, because the weights live in a public registry. After that single download, you can disconnect permanently and the AI still works. For reference material that also needs to be available offline, see using AI offline and the knowledge-packs feature covered in knowledge packs explained.

What AI models can I run locally?

The three families InnerZero installs by default are Qwen 3, Gemma 3, and gpt-oss. Together they cover everything from a 1B voice model on a basic laptop up to a 235B reasoning model on a datacenter-class workstation.

Qwen 3 (by Alibaba) is the daily-driver family: Apache 2.0 licensed, sizes from 4B to 235B, strong at chat and code. Gemma 3 (by Google) covers the smaller sizes well and powers voice transcription. gpt-oss (OpenAI's open-weight release, 120B) is the top end of the default install on high-memory machines. For the complete catalogue and hardware mapping, see the models page.

What can I do with local AI?

The same tasks people use ChatGPT for, minus the cloud round trip: writing, coding, summarising, planning, answering questions about your own documents, voice chat, image understanding.

The practical surface area looks like this:

• Chat: day-to-day conversation, idea sparring, writing help.
• Code: reading, writing, debugging, explaining across common languages.
• Voice: push to talk, wake words, fully local speech-to-text.
• Knowledge packs: your own PDFs, notes, or reference material as offline context.
• Tools: local file access, search, and calculators where supported by the front end.

How does InnerZero fit?

InnerZero is a desktop app that turns a local runtime (Ollama or LM Studio) into a polished daily assistant. It adds memory, voice, tools, knowledge packs, and an optional cloud layer for when you want a frontier model.

The separation of concerns is intentional: Ollama and LM Studio are excellent engines but offer only a developer-oriented chat UI. InnerZero is the interface that makes local AI feel like a product instead of a tinkering project. Cloud is optional and always opt-in: add your own API key for any of seven providers, or subscribe to the managed plan. Either way, local stays free forever. Start at the features page, grab the installer from the download page, or see the full breakdown on the pricing page.

Who is local AI for?

Most people benefit from local AI, but the pitch differs by context. Pick the persona that matches your use case for a tailored breakdown.

Frequently asked questions