Open vs Closed Models

Open vs Closed Models

One of the most significant structural debates in AI right now is between open and closed model development. The choice has real consequences for developers, businesses, and society — and the landscape is shifting fast.

What "Open" Actually Means

The term "open source" in AI is often imprecise. There are several different things people mean:

Open weights: the trained model weights are publicly available. You can download and run the model. This is what Llama, Mistral, and Qwen provide. You can't necessarily retrain from scratch (training data often isn't released), but you can run, fine-tune, and deploy.

Open source (strict): weights AND training code AND training data are all released. Very rare for frontier models — mostly academic models like Pythia.

Open API: the model is accessed via an API but the weights aren't released. ChatGPT, Claude, Gemini. You can use them but not run them yourself.

Most "open source AI" in practice means open weights.

The Case for Open Models

Cost: run locally or on your own infrastructure — no per-token API charges at scale.

Privacy: data never leaves your premises. Critical for healthcare, legal, and government use cases.

Customisation: fine-tune on private data without sending it to a third party.

Reliability: no vendor dependency. No API outages, no sudden price changes, no model deprecations.

Community: hundreds of fine-tunes, tools, and adaptations built by the community.

Research and auditing: academic researchers and security teams can inspect model behaviour more deeply.

Licensing is getting more complicated: early open-weight releases mostly used permissive licences (Apache 2.0, MIT). Increasingly, labs release weights under custom licences with conditions attached — limits on commercial scale, restrictions on using the model to improve a competing model, and so on. "Open weights" no longer reliably means "unrestricted use" — check the licence before deploying commercially.

The Case for Closed Models

Capability: frontier closed models still tend to lead on the most demanding tasks — complex reasoning, coding, long-context understanding — though the gap with the best open-weight models has narrowed considerably.

Convenience: no infrastructure to manage. Call an API, pay per token, iterate.

Safety investment: frontier labs invest heavily in alignment, red-teaming, and safety training. Less risk of deploying a model that produces harmful outputs.

Multimodality: best vision, audio, and image generation capabilities are generally in closed models.

The convergence: open-weight models are catching up fast. For many tasks — code generation, summarisation, classification, translation — open-weight models already match or exceed older closed models, at far lower cost.

As of October 2026: closed frontier families include ChatGPT, Claude and Gemini. The strongest open-weight models — DeepSeek, Qwen, Kimi and GLM among them — are close behind on most benchmarks, and sometimes ahead on specific ones.