For most of the last decade, the most capable artificial-intelligence models were closed: you could use them through an interface, and you could not inspect or host them yourself. That has changed. A growing family of open-weight models now performs close to the frontier on many tasks, and the shift has consequences for cost, control and competition.
Open weights mean the model parameters are published, so anyone with the hardware can run the system and adapt it. That is a different proposition from open source in the strict software sense, because the training data and process often remain undisclosed. The practical effect is still significant.
Why the gap narrowed
Several factors converged. Research on training methods became more widely shared, compute-efficient techniques improved, and the cost of training a capable model fell. Smaller teams and public research groups began releasing models that matched earlier generations of the largest commercial systems. The frontier keeps moving, and the trailing edge has risen sharply.
What open weights enable
Organisations can run models on their own infrastructure, which matters where data cannot leave a jurisdiction or a secure facility. They can fine-tune a model on domain data without sending it to a vendor. They can inspect behaviour, audit outputs and freeze a version rather than depend on a service that may change. Each of these is a form of control that closed systems do not offer.
Running a model locally shifts spending from per-token fees to hardware and operations. For heavy users, that can be cheaper; for occasional users, it usually is not. The economics depend on utilisation, and the calculation favours organisations with steady, high-volume workloads and the engineering capacity to run them.
The safety debate
Publishing weights removes the ability to withdraw a model or restrict its use. Proponents argue that openness supports scrutiny and prevents concentration of power. Critics argue that it also removes guardrails and makes misuse harder to prevent. The disagreement is genuine, and it has not been resolved by evidence, because the field is young.
The practical decision for a company is no longer whether to use AI but where the model runs and who controls it. Open weights give an exit option that reduces dependence on a single supplier. That option has value even for organisations that choose a hosted service, because it changes the negotiating position.
Where closed models still lead
The largest closed models retain advantages in capability, tooling and support. They also handle scale and reliability that smaller deployments struggle to match. For many applications, the convenience of a managed service outweighs the control of running your own. The choice is a trade-off rather than a settled answer.
Open and closed models are likely to coexist, with the frontier shifting and the trailing edge rising. What matters for policy is that both paths remain available, so that control and capability can be balanced according to the needs of the user rather than the business model of a single provider. That balance is now a live question in several jurisdictions.
The licensing question
Open weights are often released under licences that permit commercial use with conditions. Some restrict use for certain purposes, some require attribution, and some limit redistribution. These terms matter for companies deciding whether they can build a product on a model, and they are more varied than the simple label “open” suggests. Reading the licence is part of the decision.
Several governments have funded open model development as a way to avoid dependence on a few foreign providers. Public investment in open models is a form of digital sovereignty, and it also supports domestic research capacity. Whether it produces competitive systems or simply subsidises a parallel effort is an open question.
What the benchmark numbers hide
Leaderboards compare models on standard tests, and the scores are easy to quote. They also compress very different capabilities into a single figure, and models can be tuned to perform well on the tests themselves. A model that scores highly may still fail on the specific task a business needs, which is why evaluation on your own data matters more than any public ranking.
Training and running large models consumes electricity and water. Open models let organisations choose where that load runs, which can shift it toward cleaner or dirtier grids. The environmental effect of AI depends heavily on where the compute sits, and that is a decision that open weights put in more hands.
The talent question
Running open models well requires engineers who understand inference, hardware and evaluation. That skill is scarce and expensive, which means the control that open weights promise is only fully available to organisations that can staff it. For everyone else, the practical choice is still between managed services, and the openness of the underlying model changes the negotiation rather than the day-to-day operation.
This report draws on the OECD AI Policy Observatory and the Stanford AI Index.
Image: 123net · CC BY-SA 3.0 · via Wikimedia Commons.