Few questions in AI provoke as much debate as whether powerful models should be released openly. Some of the world’s leading AI developers keep their best models closed, available only through apps and APIs. Others release model weights that anyone can download, run and modify. The choice affects innovation, competition, safety, national strategy and who controls the future of the technology.

Part 1 — What “open” means

“Open” is used loosely. It helps to distinguish:

  • Closed models: accessible only through the developer’s products or API. Users cannot see or download the weights.
  • Open-weight models: the trained parameters are published for download, usually with a license. Users can run them locally, fine-tune them and build on them. Training data and full training code are often not released.
  • Fully open models: weights, training code and training data (or detailed data descriptions) are released, allowing others to reproduce and study the model.

Licenses vary widely. Some open-weight models use permissive licenses such as Apache 2.0; others use custom licenses with restrictions on certain uses or very large companies. In 2024, the Open Source Initiative published an Open Source AI Definition, which requires the freedom to use, study, modify and share a system, including sufficient information about the training data to build a substantially equivalent system — a standard many “open” models do not meet.

Part 2 — The landscape

  • Closed frontier models are offered by companies such as OpenAI, Anthropic and Google (for its largest Gemini models).
  • Open-weight models have come from Meta (Llama), Mistral, Google (Gemma), Alibaba (Qwen), DeepSeek, Microsoft (Phi) and many others, with OpenAI also releasing open-weight models in 2025.
  • Fully open research models come from groups such as the Allen Institute for AI and EleutherAI.

Open-weight models have generally trailed the best closed models by a margin that has varied over time — sometimes narrowing to months on many benchmarks.

Part 3 — The case for openness

Innovation and access. Researchers, startups and developers worldwide can build on open models without permission or per-token fees, adapting them to new languages, domains and devices.

Transparency and research. Scientists can study model internals, test safety properties and reproduce results — essential for interpretability and safety research.

Competition. Open models prevent a few companies from controlling access to a foundational technology, and put downward pressure on prices.

Privacy and control. Organizations can run models on their own infrastructure, keeping sensitive data in-house.

Resilience and sovereignty. Countries and organizations reduce dependence on foreign or single providers.

Historical precedent. Open-source software, from operating systems to web servers, became the backbone of the internet, and is often considered more secure because many people can inspect it.

Part 4 — The case for keeping models closed

Misuse risk. Once weights are released, safeguards can be removed through fine-tuning, and the release cannot be recalled. Critics worry about help with cyberattacks, disinformation at scale, or — for future, more capable systems — assistance with biological or chemical weapons.

Monitoring and control. API access lets developers detect abuse, update safety measures and cut off malicious users.

Responsible scaling. Some developers argue that as capabilities grow, the bar for open release should rise; several have published frameworks that tie release decisions to capability evaluations.

Business sustainability. Training frontier models costs enormous sums; closed access is the main way some developers recover these costs.

Part 5 — The evidence on marginal risk

A central question is marginal risk: does releasing a model meaningfully increase risk compared with what is already possible using closed models, search engines and existing knowledge? Researchers have argued that assessments should focus on this marginal difference rather than on absolute capabilities. Evidence varies by domain and model generation, and the answer may change as capabilities increase.

Part 6 — Business strategies

Companies release open weights for strategic reasons:

  • ecosystem building: developers adopt their models and tools;
  • commoditizing complements: lowering the value of competitors’ closed models;
  • talent and reputation: attracting researchers;
  • cloud and services revenue around open models;
  • standard setting: becoming the default foundation others build on.

Closed developers, meanwhile, compete on frontier capability, product integration, reliability and enterprise trust.

Part 7 — Policy

  • The EU AI Act includes obligations for providers of general-purpose AI models, with certain exemptions for models released under free and open-source licenses — but not for models deemed to pose systemic risk.
  • In the United States, policy debates have weighed the benefits of open models against risks, generally without restricting open-weight releases so far.
  • Export controls and national security concerns increasingly shape discussions, particularly as strong open models emerge from different countries.

Part 8 — A spectrum, not a binary

Many experts now advocate a graduated approach: releasing models openly when risk evaluations show acceptable marginal risk, staged releases, structured access for researchers, and stronger safeguards as capabilities grow. Both open and closed models are likely to coexist, serving different needs.

How to choose as a user or business

  • Need maximum capability with minimal setup? Closed frontier models via API.
  • Need data control, customization or offline use? Open-weight models.
  • Need to study or reproduce models? Fully open research models.
  • Check licenses carefully for commercial use.

A practical example: choosing for a hospital

A hospital wants an assistant to summarize clinical notes. A closed frontier model via API offers the best quality and minimal setup, but patient data would leave the hospital’s infrastructure, requiring strict contracts and compliance reviews. An open-weight model can run on the hospital’s own servers, keeping data in-house and allowing fine-tuning on local terminology, but requires technical staff, hardware and its own safety testing.

Many organizations in this position test both: they evaluate quality on real (de-identified) notes, estimate total costs including staff and hardware, and assess regulatory requirements. Some choose closed models with enterprise data protections; others choose open models for control; others use open models for sensitive data and closed models for less sensitive tasks. The debate, in practice, becomes a series of concrete trade-offs.

Questions that cut through the debate

  • What capabilities does the model have, and what is the realistic marginal risk of releasing it?
  • Who benefits from openness or closure in this case — users, researchers, competitors, the developer?
  • What safeguards exist after release, and how effective are they?
  • What would we lose by keeping this model closed — research, competition, access?

What to watch next

  • The capability gap between open-weight and closed frontier models.
  • Licensing trends and adoption of the Open Source AI Definition.
  • EU AI Act obligations for open general-purpose models with systemic risk.
  • Evidence on the marginal risks of open releases.

Key takeaways

  • Closed, open-weight and fully open models differ in what is released.
  • Openness supports innovation, research, competition, privacy and sovereignty.
  • Closed access allows monitoring and control but concentrates power.
  • Debates increasingly focus on marginal risk compared with existing alternatives.
  • A graduated approach and coexistence of open and closed models are likely.

Frequently asked questions

Is an open-weight model the same as open source?

Not necessarily. Open source AI, under the OSI definition, requires more, including information about training data and freedoms to use, study, modify and share.

Are open models less safe?

They can have safeguards removed, but they also enable independent safety research. Risk depends on capabilities and the marginal difference from existing tools.

Can I use open-weight models commercially?

Often yes, but licenses vary. Some include restrictions, so check terms before building products.