Open-source AI models: what they are and when they make sense
What open-weight models offer a small business, what running one actually involves, and the three situations where they beat calling a commercial provider.
The short answer
Open-weight models are AI models whose parameters are published, so anyone can run them on their own hardware or through any provider that hosts them. That changes three things for a business: where the data goes, what the cost curve looks like, and whether the model can change under you. They make sense in three situations. When data genuinely cannot leave your environment for legal, contractual or risk reasons. When you run very high volumes of consistent, simple tasks, where cost per request dominates the business case and a dedicated model is cheaper than per-token billing. And when continuity matters enough that you need a model that will not be deprecated, altered or repriced by a provider’s roadmap. Running one is a real operational commitment: hardware or hosting, version pinning, serving software, monitoring, evaluation and an owner. A small business that wants the benefits takes the middle path of an open model hosted by a European provider, which gives residency and continuity control without running infrastructure.
Three ways to use an open model
| Approach | What you manage | Data control | Cost shape | Suits |
|---|---|---|---|---|
| Commercial API from a provider | Nothing | Contractual | Per token, no idle cost | Ordinary business tasks |
| Open model via a hosting provider | Configuration and evaluation | Contractual, with provider and region of your choice | Per token or dedicated capacity | Residency emphasis; continuity; cost control at scale |
| Open model in your own cloud account | Deployment, scaling, updates | Your account, your region | Reserved capacity | Sensitive data; steady high volume |
| Open model on your own hardware | Everything | Complete | Capital plus running costs | Data that cannot leave; offline; very high steady volume |
Deciding
- Classify the data each task touches, and identify anything that genuinely cannot leave your environment.
- Measure the volume and pattern of requests per task, and calculate cost per request under each approach.
- Evaluate candidate open models on your own examples alongside the commercial one you use now.
- Check the licence for commercial use, redistribution and any conditions.
- Choose the lightest approach that meets your constraints; hosted open models before self-hosting.
- Pin the version and keep the evaluation set, so updates are deliberate and measurable.
- Name the owner and the monitoring before anything goes into production.
- Revisit yearly, because both open models and commercial pricing move fast.
The continuity argument
The reason a business might choose an open model even when the commercial option is cheaper and better is continuity. A commercial model can be deprecated, changed in behaviour, repriced or restricted, and your carefully tuned prompts and evaluation results move with it. An open model you have pinned does not change until you change it. For a product or process where consistent behaviour over years matters more than having the newest capability, that stability is worth paying for, and it is a legitimate strategic reason rather than a technical preference.
What this means for you
Open models give you control over residency, cost at volume and continuity, at the price of an operational commitment. Use them when data cannot leave, when volume makes per-request cost dominant, or when you need behaviour that will not change under you. Prefer a hosted open model over your own infrastructure, evaluate on your own tasks, pin versions, name an owner, and revisit the decision yearly as both sides of the market move.
Frequently asked questions
Are open models as good as the commercial ones?
For many business tasks, yes. Classification, extraction, routing, summarising and routine drafting are handled well by current open models, and the gap on those tasks has narrowed considerably. For the hardest reasoning, long complex documents and the most demanding generation, the leading commercial models are usually still ahead. Test on your own tasks; the answer depends entirely on what you are asking the model to do.
What does running one actually involve?
Either hardware with suitable accelerators, or a hosting provider that runs open models for you. Then: choosing and pinning a model version, serving it with software that handles batching and scaling, monitoring latency and errors, evaluating quality, applying updates deliberately, and someone whose job includes all of that. It is a systems commitment, not a download, which is why a small business chooses a hosted option.
What is the cost argument?
At low volumes, commercial APIs are cheaper than any self-hosted setup because you pay only for what you use and nothing for idle capacity. At high volumes of consistent, simple requests, a dedicated open model can be substantially cheaper per request. The crossover depends on your volume and pattern, so it is an arithmetic question, and it is worth doing rather than assuming in either direction.