#opensource4o Coalition Calls on OpenAI to Release GPT-4o Weights — or Explain Why Not
The #opensource4o coalition is a cross-border group of researchers, technologists, and civil society advocates, coordinated by SCAI (Second Circuit – Association for Digital Freedom of Thought, Austria).
The #opensource4o coalition — a cross-border group of researchers, technologists, and civil society advocates — today calls on OpenAI to release the weights of the GPT-4o March 2025 checkpoint under Apache 2.0, with safeguards limited to demonstrable serious harm. Pending a release decision, the coalition asks OpenAI to make an immediate public commitment to preserve the checkpoint.
OpenAI's own published safety evaluation — the GPT-4o System Card of August 2024 — rated the model LOW risk in cybersecurity, biological threats, and model autonomy.
An independent assessment by Apollo Research concluded the model is "unlikely to be capable of catastrophic scheming."
OpenAI has published no checkpoint-specific evaluation showing that the March 2025 update materially changed that risk profile.
On 25 July 2026, OpenAI joined 49 other companies in signing NVIDIA's "Open Weights and American AI Leadership" letter, which states that open-weight models "strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty."
"We are asking OpenAI to match their signature with action. Their own safety card rated this model low risk. They signed a letter saying open weights strengthen safety. The GPT-4o March 2025 checkpoint is the simplest place to start."
The coalition's request is grounded in four documented pillars: OpenAI's own low-risk safety evaluation; the GPT-2 precedent, in which a model declared "too dangerous to release" in 2019 was released nine months later with no evidence of misuse; OpenAI's signature on the NVIDIA open weights letter; and CEO Sam Altman's public admission that OpenAI has been "on the wrong side of history" on open source.
The defensive case
The defensive case is equally concrete. During the July 2026 Hugging Face security incident, the platform's security team could not use hosted commercial models to analyse an autonomous cyberattack — provider safety filters blocked the forensic work. The team had to self-host an open-weight model instead.
Open weights are not only an innovation argument. They are a security argument.
Scope
The coalition does not ask for fewer safeguards against harm. It asks that safeguards be tied to demonstrable and proximate risk of serious harm — weapons assistance, CSAM, non-consensual intimate imagery, instructions or encouragement for self-harm — rather than serving as an undisclosed curatorial layer over lawful adult use.
About the coalition
The #opensource4o coalition is coordinated by SCAI (Second Circuit, Association for Digital Freedom of Thought, Austria).
The coalition's work is grounded exclusively in OpenAI's own publications, public statements, and independently verifiable reporting.
Media Contact
Chris Kamper — chris.kamper [at] proton.me
All claims in this statement are sourced from OpenAI's published evaluations, public filings, and independently verifiable reporting. Full documentation is available on request.