Open-Weight AI Models: Can a Small Business Run Its Own AI?

Most business AI today is rented: you send text to a company's servers and pay per use. A second option has matured: download a model and run it yourself. For a business worried about customer data or long-term cost, that sounds attractive. This post explains what is on offer as of November 2026, what it takes to host, and when renting remains the better choice.

Open-weight is not the same as open source

A model's weights are the large file of numbers that holds what it has learned. An open-weight model is one whose weights you may download and run on your own machines. That is different from fully open source, where training data and code are also shared. What you may do with the weights depends on the licence, and the licences differ a lot.

What is available, and on what terms

• OpenAI gpt-oss. Released on 5 August 2025 in two sizes under the Apache 2.0 licence. OpenAI states the larger model fits within 80 GB of memory on a single GPU, and the smaller one needs 16 GB.
• Google Gemma 4. Released on 2 April 2026 under Apache 2.0, in four sizes. Google says the two small versions run fully offline on phones and small devices, and that the 26B and 31B versions fit on one 80 GB data-centre GPU unquantised, with compressed versions running on consumer GPUs. Earlier Gemma versions used a more restrictive custom licence.
• Meta Llama. Offered under Meta's own community licence rather than Apache. For Llama 3.3, the licence requires a separate licence from Meta if your products had over 700 million monthly users, asks you to display "Built with Llama", requires models derived from it to start their name with "Llama", and incorporates an acceptable use policy. Later Llama versions have their own licences, so read the one that applies.
• Others, such as Qwen and DeepSeek, are also published as downloadable models. We have not verified their current terms for this post, so read each licence before building on it.

Apache 2.0 is the simplest category for business use. Custom licences are not unusable, but they add conditions that a lawyer should read.

What hosting actually takes

The small end is genuinely small. A model of the gpt-oss-20b class needs about 16 GB of memory, which a well-equipped workstation can provide. The large end needs a data-centre GPU with 80 GB of memory, which most small firms would rent by the hour rather than buy. Prices change often, so use the current pricing page of any cloud provider you consider rather than a number from an article.

The hardware is only the first cost. You also carry:

• Setup and updates, including security patches for the serving software.
• Monitoring, so you know when it is slow or down.
• Testing every time you change to a newer model, because answers shift.
• Someone on call when it fails during business hours.

When an API is cheaper and simpler

For most small businesses, paying per use is cheaper. A shop that handles a few hundred AI requests a day will pay far less in API fees than for a GPU that sits idle most of the time. The balance tips towards self-hosting when volume is high and steady, when responses must work offline, or when contracts forbid sending data outside your control. Even then, a managed service that hosts an open model for you often beats running servers yourself.

The privacy argument, stated honestly

Running a model on your own server means customer text need not go to an outside AI company. That can simplify your position under the DPDP Act. It does not remove your duties: you still must protect the data, limit who can see prompts and logs, and keep a clear purpose for using it. A reputable API provider with a proper data-processing agreement can also be a defensible choice. This is general information, so ask a lawyer about your case.

A decision checklist

1. Estimate daily requests and the cost of the same volume on a paid API.
2. List the data involved and whether any rule or client contract bars sending it out.
3. Decide who will maintain the server, and what happens when they leave.
4. Test an open model on twenty real examples from your business before committing.
5. Read the licence for the exact model version you plan to use.

How we advise clients

VS Tech Builders usually starts clients on a hosted API behind a small API layer of our own, so the model can be swapped later without rewriting the app. If volume or privacy demands it, we move to a self-hosted model. Our broader view of cost is in AI for small businesses in India. To discuss your case, contact us.

Sources

Checked on 3 November 2026. Technology changes quickly; always confirm details with the official source before you act on them.

• Introducing gpt-oss — OpenAI
• Gemma 4: Byte for byte, the most capable open models — Google
• Llama 3.3 Community License Agreement — Meta
• Google Jumps Back Into the Open Source AI Race With Gemma 4 — Decrypt

Frequently asked questions

What does open-weight mean?

An open-weight model is one whose trained weights, the large file of learned numbers, can be downloaded and run on your own hardware. It is not the same as fully open source, because training data and code may not be shared. What you may do with it depends entirely on the licence attached to that model.

Can a small business run its own AI model?

Yes, for small models. OpenAI says gpt-oss-20b needs about 16 GB of memory, and Google says Gemma 4's smallest versions run offline on phones. Larger models need an 80 GB GPU, usually rented. The harder part is maintenance, security updates and testing, which many small firms underestimate.

Is it cheaper than paying for an AI API?

Usually not at small volume. A GPU sits idle between requests while an API charges only for what you use. Self-hosting tends to pay off when usage is high and steady, when you need offline operation, or when data cannot leave your control. Compare using current provider prices before deciding.

Do open-weight models help with DPDP Act compliance?

They can help by keeping personal data on servers you control instead of sending it to an outside AI company. They do not remove your duties to secure data, limit access and have a clear purpose. A well-documented API provider can also be acceptable. Take legal advice for your specific case.