Small AI Models Are Moving Onto Your Own Computer: What It Could Mean for Small Businesses

What launched this week?

Two launches on 27 October point the same way: AI that runs on the computer in front of you, not on someone else's server.

• Liquid AI released LFM3. Liquid's announcement describes a family of open-weight models, with text, vision and audio versions, in the 2-billion-parameter range. You can download them from Hugging Face. The small text model, LFM3-1.7B, takes about 1 GB of memory at 4-bit precision, and Liquid says its models can run on any device.
• Opera began shipping local AI in its browsers. Opera's announcement says an early preview of WebAI is in Opera One, with Opera GX and Opera Neon to follow. It is opt-in, and nothing is downloaded until you turn it on. At launch it offers models from Google and Apple, including Google's Gemma and, on Macs, Apple's on-device model.

Neither one is a product you need to go and buy. Both are early.

Why could it matter to a small business?

Think of it as three possible shifts. These are directions, not promises.

• Small jobs may stop needing a subscription. Opera's pitch is that users "don't need AI subscriptions" for jobs a laptop can handle. Rewording an email or sorting a list is the sort of thing that might move over. AgentShed hasn't measured whether it saves anyone money.
• Your existing tools may add it quietly. A browser, a phone or an office app could start using a local model behind a button you already know. You wouldn't have to set anything up.
• More choice. Opera says developers can add AI features without running AI infrastructure or paying for every request. If that holds, more small tools could appear, and you would have more to choose from.

What can't small models do yet?

This is where the hype runs ahead of the facts.

• Liquid's own model card says the 1.7B model is designed for local agents, tool calling and fine-tuning for a specific field, and is not recommended for agentic coding or complex multi-step reasoning because of its small size.
• Opera says performance depends on your hardware, and the preview covers 64-bit Windows PCs and Apple-silicon Macs only.
• The strongest models still live in data centres. NVIDIA's open Nemotron 3.5 Transcribe, released 28 October, transcribes recordings of up to an hour in 40 languages and can label who is speaking. Its model card lists one 80 GB H100 data-centre GPU for the standard set-up, and NVIDIA and Hugging Face's launch post points to hosted endpoints. For most people that means using it through a hosted service, not a laptop.
• Even Google's newest frontier model, Gemini 4 Carbon, is going first to a subset of cyber defenders, with access phased in for trusted partners. Google says it will reach developers, businesses and consumers "as soon as possible".

Get the next guide by email. One email when we publish a new AI workflow. Join the AgentShed newsletter.

Which route suits you today?

Prices and plan limits change quickly, so we haven't put any in this table. Check each vendor's page on the day.

What should you do now?

You don't need to buy anything or install anything. A sensible next step is small:

1. Keep using hosted free plans for real work. Start with one task, such as drafting a reply, with a made-up example. Our customer FAQ replies playbook and social repurposing playbook are good first tries.
2. Write down three jobs that are simple and repeat. Rewording, sorting and renaming are small, simple jobs, and a sensible place to start. Use invented practice text, never real customer details, if you test any of them.
3. Watch your own tools. If your browser, phone or office suite adds a local option, try it on one low-risk job and compare the result with what you do now.
4. Come back in a few months. This area moves fast. Our October 2026 trends page and the newsletter will flag when something is ready for a small business.

How we made this

AI-assisted, and fact-checked against the vendors' own pages on 29 October 2026. We haven't downloaded, run or tested LFM3, Opera's WebAI or Nemotron 3.5 Transcribe, so nothing here is a tested result. Where a claim is the vendor's, we say so. See how we research a playbook.

FAQ

Do I need to run an AI model on my own computer? No. Hosted apps remain the simplest way to start, and most guides on this site use them. Local models are an option to watch, not a requirement.

Can a small business use LFM3 for free? Liquid's licence summary says free commercial use applies while a company's annual revenue is under $10 million, and that you should contact Liquid for a commercial licence once you pass it. The licence text defines the threshold as $10,000,000 or more and counts companies you control or share control with. That is Liquid's own wording relayed here, not legal advice, so read the licence page and ask Liquid if you are unsure.

Will running AI locally be cheaper than a chat subscription? We don't know, and AgentShed hasn't measured it. Opera says developers can add AI features without paying for every request, but you would be using your own computer and your own time to set it up. Any saving depends on the job.

Is Opera's local AI ready for everyone? Not yet. Opera calls it an early preview, it is opt-in, and it covers 64-bit Windows PCs and Apple-silicon Macs only. Expect rough edges.

Your next step

Pick one repeat job this week, write a made-up example of it, and try it in a hosted free plan. That one habit will help you judge any new tool, local or not.