Open-weights models are machine learning models, typically large language models or other foundation models, whose trained parameters, or weights, are released publicly for anyone to download, run locally, inspect, and fine-tune. This contrasts with proprietary or closed models, such as GPT-4 or Claude, which are accessible only through a paid API or hosted interface, with the underlying weights kept private.
Terminology
Open-weights is often used instead of open-source because releasing weights alone does not meet traditional open-source definitions, which typically require the training data, training code, and methodology also be public and freely reusable. Most widely used open models, including Llama, Mistral's releases, and Qwen, publish weights and often architecture details, but not the full training data or complete training recipe, making open-weights a more precise term than open-source AI, a phrase organizations like the Open Source Initiative have worked to define more strictly.
History
Early influential open releases include Meta's Llama models, first released in February 2023 and initially leaked before an official broader release, which sparked a wave of community fine-tuning and derivative models. Stability AI's Stable Diffusion (2022) played a similar role for image generation. Through 2023-2025, open-weights releases proliferated, including Mistral AI's models, Alibaba's Qwen family, and, notably, DeepSeek's DeepSeek-R1 in January 2025, an open-weights reasoning model whose training efficiency claims contributed to the DeepSeek market shock. Nonprofit and research efforts, including the Allen Institute for AI's OLMo project, have pursued fuller openness, releasing training data and code alongside weights.
Licensing
Open-weights models are released under varied licenses, from permissive licenses close to traditional open source, to custom licenses that impose conditions, such as usage caps tied to a company's revenue or restrictions on using outputs to train competing models. This variation has made "open" a contested label, since two models both described as open-weights may carry meaningfully different legal terms.
Open versus closed debate
Proponents of open-weights release argue it democratizes access to powerful AI, enables independent safety research and interpretability work, avoids vendor lock-in, and allows deployment in privacy-sensitive or offline settings. Critics, including some voices within AI safety, argue that once weights are public, safety mitigations built into a hosted model, such as guardrails against harmful use, can be removed by anyone with the technical skill to fine-tune the model, and that open release removes any ability to revoke access to a dangerous capability after the fact. Figures such as Yann LeCun have argued for openness as essential to competition and scientific progress, while others favor more cautious, staged release practices, as OpenAI initially adopted with GPT-2 in 2019.
Economic and geopolitical dimensions
Open-weights releases have become a competitive strategy: Meta positioned Llama as a way to commoditize the base-model layer while it competes on products and infrastructure, and Chinese labs' aggressive open releases have been read partly as a response to US export controls on advanced chips, since a widely adopted open model extends influence even where access to leading proprietary APIs is restricted. As of 2025, open-weights models from several labs approach the performance of leading closed models on many benchmarks, though the very largest frontier models have generally remained closed.