This editorial sets out the publication’s position.

Few debates in AI generate as much heat as the one between “open” and “closed” models. Advocates of openness warn that a few companies could control a technology as fundamental as electricity. Advocates of closed release warn that publishing powerful models’ weights makes misuse impossible to stop. Both sides make valid points. The debate goes wrong when it becomes a binary choice.

What openness provides

Open-weight models — whose parameters anyone can download — have delivered clear benefits:

  • Research and transparency: independent scientists can study how models work, test their safety and reproduce results.
  • Competition and lower prices: open alternatives constrain the market power of leading providers.
  • Access and sovereignty: universities, startups and countries without massive budgets can build on capable models, including for their own languages.
  • Privacy: organizations can run models on their own infrastructure, keeping sensitive data in-house.
  • Innovation: a global community fine-tunes, compresses and adapts open models for uses their creators never imagined.

What closed release provides

Keeping models behind an API also has real advantages:

  • Ongoing safeguards: providers can monitor misuse, update filters and revoke access.
  • Staged release: capabilities can be tested with limited audiences before wider availability.
  • Accountability: a single provider is responsible for the system’s behavior.
  • Sustainable business models that fund expensive research.

Why the binary framing fails

  1. Openness is a spectrum. Models can be released with weights but not training data, with restrictive licenses, or to vetted researchers only. “Open-weight” and “open-source” are not the same thing.
  2. Risk depends on capability. Releasing the weights of a model with modest capabilities poses very different risks than releasing one that could meaningfully help create weapons. Policy should scale with evidence about capabilities, not with labels.
  3. Both can fail. Closed models can be jailbroken; open models can be fine-tuned to remove safeguards. Neither approach guarantees safety.
  4. The ecosystem needs both. Many real-world systems combine a closed frontier model with open models for specialized or private tasks.

What good policy looks like

  • Evaluate capabilities before release, for open and closed models alike, focusing on severe risks.
  • Avoid blanket bans on open weights that would entrench incumbents without clear safety benefits for most models.
  • Support open research infrastructure, including datasets and compute for academia.
  • Be precise with terms — and honest about trade-offs.

Our position

We believe a healthy AI ecosystem includes both open and closed models, governed by rules proportionate to demonstrated risk. The goal is not to pick a side, but to keep the benefits of openness while managing the genuine risks of the most capable systems.

What we will watch

We will keep reporting on the evidence: how open and closed models compare on capabilities and safety evaluations, how open releases are used and misused in practice, and how regulation treats each. Our position may evolve with that evidence — which is how it should be.

For the full arguments behind this editorial, read our deep dive on open vs. closed AI models, our explainer on frontier AI safety frameworks and our guide to open-source AI projects every developer should know. Each covers evidence that shapes our view.

Bottom line

Open and closed models serve different needs, and a healthy AI ecosystem needs both — with safeguards scaled to real, measured risks rather than to fears or slogans.

What to watch next

  • Regulatory treatment of open-weight models in the EU, US and elsewhere.
  • Capability evaluations published before major open-weight releases.
  • How the gap between open and closed frontier models evolves.

Key takeaways

  • Open-weight models bring transparency, competition, access, privacy and innovation.
  • Closed release allows ongoing safeguards, staged release and accountability.
  • Openness is a spectrum, and risk depends on capability, not labels.
  • Neither approach guarantees safety; the ecosystem benefits from both.
  • Policy should be proportionate to evidence about capabilities.

Frequently asked questions

Is open-weight the same as open-source?

No. Open-weight models publish trained parameters; open-source AI, as defined by the Open Source Initiative, also requires sufficient information about training data and code.

Are open models more dangerous?

It depends on their capabilities. For most models the benefits are substantial and risks modest; the most capable systems warrant careful evaluation before any release.

Does The Latent Daily favor one side?

No. We cover both and argue for policies proportionate to evidence.