What it is
“Open-weight” describes an AI model whose trained parameters — the numerical values, or “weights,” that a neural network learns during training — are published for anyone to download, run, and fine-tune. Meta’s Llama family, Mistral’s models, Alibaba’s Qwen series, and DeepSeek’s R1 are commonly cited examples. Users can inspect the model’s behavior, adapt it to their own data, and deploy it on their own infrastructure without depending on a vendor’s API.
What distinguishes open-weight from “open source” is what is withheld. The Open Source Initiative (OSI), which published version 1.0 of its Open Source AI Definition in October 2024 after a two-year, multi-stakeholder drafting process, holds that a system only qualifies as open source AI if the weights, the code used to train and run it, and the training data — or “sufficiently detailed information” about that data — are all released under terms granting the freedom to use, study, modify, and share the system for any purpose. Most open-weight releases fall short of that bar: they publish the weights but not the training corpus, and they attach custom licenses rather than recognized open-source ones.
Meta’s Llama license illustrates the gap. OSI has stated plainly that the Llama license “is still not Open Source.” The license imposes an acceptable-use policy restricting certain applications, requires a separate commercial license from companies with more than 700 million monthly active users (granted or denied at Meta’s discretion), and in recent versions has excluded users in the European Union from parts of the license entirely. The weights are open for download; the legal and technical freedoms that define open source are not.
Why it matters for AI governance and narratives
The open-weight/open-source distinction has become load-bearing in AI policy debates because it maps onto competing theories of risk and competitive advantage. One camp — including much of the open-source software community and researchers cited in the 2026 International AI Safety Report — argues that publishing weights broadens access, accelerates independent safety research, and prevents capability concentration in a handful of firms. A second camp, more prominent in U.S. national-security discussions, treats open-weight releases as a proliferation risk: once weights are public, safety guardrails can be stripped out, and the model cannot be recalled. That argument gained force after DeepSeek’s January 2025 release of R1, an open-weight reasoning model that, according to analysis from RAND and the International Institute for Strategic Studies, matched much of the performance of leading closed U.S. systems despite operating under chip export restrictions — evidence, to restriction advocates, that openness had let a strategic competitor close the gap cheaply.
This is also why open-weight has become, as one framing puts it, a supply-chain position rather than a licensing choice. A model’s weights sit downstream of the compute, power, and data infrastructure used to train it, and upstream of the applications and platforms that redistribute it. Whoever controls the licensing terms on an open-weight model controls a checkpoint in that chain even after giving away the weights themselves — which is why firms with different competitive positions (incumbent labs, challenger labs, national champions) argue for different definitions of what “open” should require.
Key facts and dates
The OSI’s Open Source AI Definition reached version 1.0 in October 2024, endorsed at launch by more than 20 organizations including the Mozilla Foundation and EleutherAI. Meta has marketed Llama as “open source” since its 2023 release despite OSI’s objection, a labeling dispute that recurred with each subsequent Llama version. DeepSeek’s R1 and R1-Zero, released in January 2025, became the most consequential open-weight release from a Chinese developer to date, prompting U.S. policy responses on export controls and renewed congressional attention to whether — and how — open-weight releases from adversary states should be restricted. Research continues on frameworks for grading openness more granularly than a binary open/closed label; the Model Openness Framework (arXiv, 2024) is one attempt to score releases along a spectrum based on which components — weights, code, data, documentation — are actually disclosed.