What it is
“Openness” in AI describes how much of a model a developer makes publicly available, and to whom. The term is used loosely in coverage of the field, but it actually spans a spectrum of distinct releases. At one end sit fully closed, API-only systems that users can query but never inspect or download. At the other end sit systems released with their weights, training code, and enough information about training data that an independent party could rebuild something functionally equivalent.
Most of what the industry calls “open” AI — Meta’s Llama family, Alibaba’s Qwen, DeepSeek’s models, Mistral’s releases — is more precisely “open-weight.” The trained parameters (the weights) are downloadable, which lets anyone run, fine-tune, or deploy the model on their own hardware without depending on a vendor’s hosted service. But the training code and the data used to produce those weights usually stay private. That distinction matters because it determines whether a model can genuinely be audited, reproduced, or modified at the source, or merely operated.
The Open Source Initiative — the nonprofit that has defined “open source” for conventional software since the late 1990s — published a formal Open Source AI Definition (OSAID) 1.0 to draw a firmer line. It requires three things for a system to earn the label: the complete source code used to train and run it, model weights and parameters released under approved terms, and “sufficiently detailed information about the data” that a skilled person could use to build a substantially equivalent system. By that standard, nearly all of the models the press and industry call “open” fail to qualify, because they withhold training data details, training code, or both.
Why it matters for AI governance and narratives
The word “openness” is doing a lot of rhetorical work in the AI debate, and different actors use it to mean different things on purpose. Companies releasing open-weight models invoke “openness” to signal democratization, scientific transparency, and resistance to concentration among a few frontier labs — a framing aimed at regulators, researchers, and developers wary of vendor lock-in. Critics counter that open-weight release is a one-way door: once weights are public, safety guardrails can be stripped with modest effort, and the release cannot be recalled if a model turns out to enable serious harm. Both claims can be true of the same release, which is why “openness” functions as a contested term rather than a settled fact — exactly the kind of framing contest this observatory tracks.
This dispute has direct policy stakes. The EU AI Act and various national frameworks increasingly tie compliance obligations to how a model is released, so whether a given system counts as “open-weight,” “open-source,” or neither can change what law applies to its developer. Proposals circulating in 2025–26 argue against treating openness as binary and instead favor tiered release — full openness for lower-risk models, restricted or sandboxed access for models whose capabilities could enable chemical, biological, radiological, or nuclear (CBRN) misuse — with access calibrated to demonstrated risk rather than to a label.
Key facts and dates
The Open Source Initiative’s Open Source AI Definition reached version 1.0 in late 2024, after a multi-year drafting process that drew on the pre-existing Free Software Definition and OECD AI principles. It set four core freedoms a system must grant — to use, study, modify, and share it for any purpose — and specified that weights, code, and training-data information must each meet that bar, not just the weights alone.
The practical gap between the OSAID standard and common industry practice remains wide. Meta’s Llama models, often described as “open source” in coverage, are licensed with restrictions (including limits on military and large-commercial use) that the OSI’s definition would not recognize as open source; a widely reported case from November 2024 described Chinese researchers repurposing a Llama model for military research despite such restrictions, illustrating the enforcement gap between a license’s terms and what is technically achievable once weights are public. DeepSeek’s R1 model, updated in May 2025, became a reference case for the risk side of the argument: independent evaluators found it matched or exceeded some frontier closed models on capability while being released with comparatively little safety evaluation. The UK AI Security Institute and others have since pushed “safety case” frameworks intended to govern release decisions for high-capability open-weight models, and OpenAI’s own July 2025 open-weight release was read by observers as early evidence of labs testing models against such frameworks before publishing them.
Where to learn more
- Open Source AI Definition 1.0 — Open Source Initiative — the primary, formally ratified definition against which “open-weight” releases are now measured.
- Beyond the Binary: A Nuanced Path for Open-Weight Advanced AI — Centre for Future Generations — policy analysis arguing for tiered, risk-calibrated release rather than an open/closed binary, with the DeepSeek R1 and Llama cases.
- Open weights are not open source: why AI’s favorite label is under dispute — The Register — accessible overview of the open-weight/open-source terminology dispute as of September 2026.
- Open AI Models: An Introduction — Simon Institute — primer on the openness spectrum across weights, data, and code.