What it is
An open-weight model is an AI system whose final trained parameters — the numerical “weights” a neural network learns during training — are published for anyone to download, run locally, and adapt. Once downloaded, users can fine-tune the model, inspect its outputs, or deploy it without depending on the company that built it or sending data to that company’s servers. Meta’s Llama family, Mistral’s models, and DeepSeek’s releases are the best-known examples (Stanford HAI).
Crucially, “open-weight” is not the same as “open source.” The Open Source Initiative — the organization that maintains the canonical definition of open-source software — argues that true open-source AI requires three things released together: the weights, the code used to train and run the model, and enough information about the training data to let others understand or reproduce it. Open-weight releases typically supply only the first of these. The training code, and especially the training data, usually remain proprietary, for competitive or legal reasons. OSI calls this a “lesser evil” compared to fully closed models, but insufficient on its own for reproducibility, bias auditing, or high-stakes deployment (opensource.org).
The distinction matters because “open” is doing a lot of rhetorical work with very little technical specification behind it. A model can be marketed as open while carrying restrictive commercial-use clauses, user caps, or export conditions — and still be described, accurately, as “open-weight.”
Why it matters for AI governance and narratives
Open-weight status has become a live regulatory category, not just a technical one. The EU AI Act exempts free and open-source general-purpose AI models from most transparency obligations under Article 53(2) — but only if the weights, architecture, and usage information are all public and the model isn’t monetized; models judged to pose “systemic risk” lose the exemption regardless. The exemption creates a concrete incentive for firms to position their releases as open enough to qualify, which is itself a framing contest playing out in how companies describe their own products.
The concept sits at the center of exactly the kind of interest-laundering the observatory watches for. The White House’s July 2025 AI Action Plan made supporting open-weight models official U.S. policy, arguing they serve startups, researchers, and government agencies with sensitive data who cannot use closed vendors. A year later, on July 24, 2026, Nvidia’s Jensen Huang used his first-ever post on X to publish a letter — signed by roughly 25 companies including Microsoft, IBM, Palantir, and CrowdStrike — arguing that keeping AI models open strengthens American safety, sovereignty, and competitiveness. Notably, OpenAI and Anthropic did not sign; both build closed, proprietary frontier models and have commercial reasons to prefer a market where openness is not the default expectation. Nvidia, whose business is selling the compute that both open and closed models run on, benefits either way — but especially from a world where open models proliferate and widen the customer base for chips, which is precisely the tension the observatory’s editorial was pointing to.
Key facts and dates
The Open Source Initiative published its formal Open Source AI Definition (OSAID 1.0) on October 28, 2024, explicitly to distinguish genuine open-source AI from the looser “open-weight” label that had proliferated in industry marketing. The EU AI Act’s general-purpose-model obligations, including the open-source carve-out, took effect as part of the Act’s phased rollout beginning in 2025. The U.S. AI Action Plan, “Winning the Race: America’s AI Action Plan,” was released July 23, 2025, and directed agencies including NIST and the NTIA to promote open-weight and open-source model adoption among small businesses, researchers, and government users. China’s DeepSeek, whose open-weight releases triggered substantial market and policy reaction in early 2025, has become a recurring reference point in U.S. debates over whether openness helps or undermines American competitiveness.
Where to learn more
- Stanford HAI: What is an Open-Weight Model? — plain-language definition and distinction from related terms
- Open Source Initiative: Open Weights — the standard-setting body’s account of why open weights fall short of open source
- Open Source Initiative: White House AI Action Plan includes open source — policy analysis of the July 2025 U.S. plan
- The National: Nvidia, Microsoft and IBM among companies endorsing open-weight AI models — coverage of the July 2026 industry letter and its non-signatories