Open-source vs closed AI models: when each makes sense for an SME
Closed models like ChatGPT and Claude are the obvious starting point. Open-source models like Llama and Mistral are quietly closing the gap. When does each make sense for a small business?
Most North Wales business owners I talk to use closed AI - ChatGPT, Claude, Gemini, Copilot. They send text to a server in California, the answer comes back, and they get on with their day. That is fine and will continue to be fine for most SMEs.
But "open-source AI" comes up in conversations more and more, usually with the question "should I be looking at this?". This article gives the honest answer.
The two worlds
Closed models are owned by a single company and accessed through their API or website. Examples: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Copilot (Microsoft). You do not see the model itself - you see what it produces. The provider hosts it, runs it, and charges for it.
Open-source models are released by a company or research group with the model files made available for anyone to download. Examples: Llama (Meta), Mistral (Mistral AI), Qwen (Alibaba), Phi (Microsoft). You can run them yourself on your own hardware, or use a hosted service that runs them for you.
"Open-source" is a slight misnomer here. Most "open" models do not release the training data, only the model weights. But the practical effect is similar: you can run the model wherever you like.
The capability gap
In 2026, the strongest open-source models are very close to the strongest closed models on most everyday business tasks. The gap that existed in 2022-2023 has mostly closed for general writing, summarising, drafting and translation work.
On the cutting edge - the most complex reasoning, the longest context, the most demanding multi-step tasks - closed models are still ahead. If your business needs the absolute best at any cost, closed wins. If your business needs good-enough at low cost, open is competitive.
Cost
This is where the two diverge most.
Closed model cost is per-use, paid to the provider. ChatGPT Plus is £18-20 a month per seat. Claude Pro similar. API costs are pence per thousand tokens, which adds up if you build a heavy automation but is small for normal interactive use.
Open-source cost is the hardware to run it (or the cost of a hosted service that runs it). Running a competitive model locally requires a modest GPU - £1,500 to £5,000 for the kind of hardware that handles a small team's daily use. Hosted open-source services (Together AI, Groq, Replicate) charge per-use rates similar to closed model APIs.
For a small business doing interactive work with one or two paid AI seats, closed is cheaper. For a business running heavy automated workloads at scale, the calculus shifts towards open-source.
Privacy and control
This is the strongest argument for open-source for some SMEs.
If you run an open-source model on your own hardware, the data never leaves your network. No third-party processor. No DPA to negotiate. No risk of a provider changing their terms. For businesses with strict data residency or confidentiality requirements - some legal, some healthcare-adjacent, some defence-adjacent - this is the deciding factor.
But "on your own hardware" requires technical capability. Most North Wales SMEs do not have a developer who can stand up and maintain a self-hosted model. Hosted open-source services give you the model choice without the operational burden, but you are back to a third-party processor.
When open-source actually makes sense for an SME
Three patterns where open-source is worth considering.
1. Heavy automated processing. If you are processing thousands of documents a day, automated content generation at scale, or running an AI feature in your own product, the per-use costs of closed APIs add up. Open-source on hosted infrastructure (or your own) can be a fraction of the cost.
2. Strict data sensitivity. If you cannot send data to a third-party even with a DPA - some regulated work, some sensitive client matters - self-hosted open-source is the only AI option that keeps data under your roof.
3. Custom fine-tuning. Some businesses train an open-source model on their own data to make it specifically good at their domain. This is more advanced than most SMEs need but is genuinely valuable in some niches.
When closed is the right answer
For the typical North Wales SME today, closed wins on most criteria.
You get the strongest models, no operational burden, paid business tiers with DPAs that satisfy GDPR for most work, predictable monthly costs, and a familiar user experience that staff can pick up without training. The tool comparison covers which closed option fits which kind of business.
Open-source becomes interesting when you outgrow the closed pattern, not as the default starting point.
The honest summary
If you are starting your AI journey, start closed. ChatGPT Plus, Claude Pro, Microsoft Copilot - these are the right tools for the first 12-18 months of most SME AI adoption.
Watch the open-source space, particularly if you process a lot of data automatically or have strict data-residency requirements. Revisit the question every six months. The space moves fast.
If you would like to think through where your business sits on this spectrum, that is what a discovery call can help with. The data privacy guide covers the closed-tier privacy detail.