What open-source models are (and when to use them)
Every organisation is now making a quiet infrastructure decision. You can rent intelligence from a vendor through an API, or you can own it and run it on hardware you control. The second option is now good enough for most work, and it is called an open-source model.
What an open-source model actually is
A closed model, the kind behind most consumer chat tools, lives on someone else’s servers. You send it text, it sends text back, and you pay for each interaction. You never see the inner workings, and you cannot move the system elsewhere.
An open-source model publishes its model files, the learned parameters that define how it behaves, so anyone can download, run and change it. Names you will hear are Llama, Mistral, DeepSeek, Qwen and Google’s Gemma. Run it on your own machine, your own server, or a small box in a classroom, and it answers without asking a foreign company for permission.
Why this matters now
For years the argument was that open models were weaker. That gap has closed. In 2025 the open-weight releases included Meta’s Llama 4, DeepSeek’s V3 and R1, Alibaba’s Qwen3, Google’s Gemma 3 and Mistral’s open models, and independent estimates put their running cost at roughly one-eighth of the equivalent closed API. The performance difference is a few months, not a different league.
- Cost. Pay for compute once, not for every question. No per-token billing.
- Privacy. Data stays on your own infrastructure. Nothing leaves to a third party.
- Sovereignty. Governments and companies control the full stack and cut foreign dependency.
- Offline use. The model runs on local hardware, air-gapped or on poor connectivity.
- Transparency. The system can be inspected, audited and tested, unlike a closed API.
- No lock-in. Retrain on your own data and move hosts whenever you choose.
Who needs to know about this
Developers should know because they can self-host, retrain on their own data and drop their API bills to near zero. Small businesses should know because it turns an unpredictable subscription into a fixed, ownable cost and keeps customer data on their premises.
Governments should know because the direction of public procurement moved toward open source through 2025. The EU’s AI Act took on general-purpose AI obligations in August 2025, and the policy conversation has since pushed toward an open-source preference for public cloud and AI buying.
Schools and non-profits should know most of all. An open model is the only kind that can run offline, for free, inside a classroom that has no reliable internet. That is precisely the gap offline-first open models exist to close.
If the tool cannot run without the internet and a foreign account, it was not built for the places that need it most. (Lawrence Oladeji)
Be precise with the language
Most models marketed as “open source” are more exactly open-weight: the model files are published, but the full training data and training code often are not. The Open Source Initiative has publicly pushed back on calling some of them fully open source. The distinction matters. When we say open here, we mean you can actually run it, inspect it and own it.
Putting this to work
An offline tutor built on this principle carries open models and local curriculum content on a small box, runs with zero internet, and serves a whole classroom. No account, no data leaving the school, no monthly bill. Open models are not a philosophy. They are what makes that possible where it is needed.
The open model is no longer the compromise. For anyone building for the edge of the network, it is the default. Own the engine.