- Open-source AI models expose more of the model stack — including code, architecture and documentation – than open-weight models
- In general, open models can help lower costs and support customization
- Understanding the open source vs. open weight distinction is key to managing lock-in, governance and deployment risk
Open-weight AI vs. open-source AI: What’s the difference? What do these terms mean?
OpenAI’s ChatGPT, Google’s Gemini and Anthropic’s Claude are what is known as proprietary (or closed-source) AI models. This means that their code and settings remain private, and the use of these models is usually paid for and strictly governed by licensing terms.
Open-source AI models, in contrast, make their structural elements available to the public and can be used freely, as IBM has noted. In 2024, the Open Source Initiative defined open-source AI models as those that can be used for any purpose without permission, allow the study of the system’s components and how it works, permit the modification of model for any purpose, and allow the system to be shared for any purpose.
Advocates have argued open-source AI models can help lower costs, enable fine tuning and provide needed transparency for model users. DeepSeek and Alibaba’s Qwen are well-known open-source AI models. IBM and Meta claim their Granite and Llama models, respectively, are open source, but there is some debate about whether they are truly open.
Comparatively, open-weight AI models release the trained model weights for download and fine-tuning, but do not disclose all training data, code or development details. Users can run and customize the model, but cannot fully inspect or reproduce it.
What are model weights?
Model weights are key settings that control how a model processes inputs into outputs. Think of as the secret sauce that determines how a model weighs (or determines) the importance of different inputs to make decisions.
Here’s a great example from Engine Advocacy Policy Fellow Min Jun Jung: “If an AI model is being trained to differentiate between dogs and cats, the model will place more importance on relevant distinguishing features (like the shape of the ears or length of nose) and less importance on less relevant features (like the number of legs or color of fur).”
Thus, an open-weight AI model makes these numerical settings publicly available.
How are open-source AI models different from open-weight AI models?
Open-weight and open-source AI models are similar in some ways, but not all open-weight AI models are open source.
An open-weight AI model can be usable and customizable without being fully open source. That distinction matters because openness is not just about access to the finished model — it’s also about transparency, reproducibility, governance and what users are legally allowed to do with it.
Open-source models generally imply that more of the model’s stack is publicly available — not just the trained model weights, but often the code, architecture details, training methodology, documentation and sometimes datasets or tooling. The idea is that others can inspect, modify, run, improve and redistribute the model under defined license terms.
What are the risks of open-source AI models vs. open-weight AI models?
Open-source AI models and open-weight AI models each present distinct risks.
Open-source models provide full access to the code, architecture, training methods and model weights, which increases transparency and auditability but also makes it easier for malicious actors to remove safety controls, modify models for harmful purposes, or create uncontrolled forks.
Open-weight models, by contrast, typically release only the trained weights while withholding training data, code and development details. This reduces transparency and can make it difficult for enterprises to assess bias, security vulnerabilities, regulatory compliance and data provenance.
In short, the primary risk of open-source AI is that it can be more easily misused, while the primary risk of open-weight AI is that organizations may not have enough visibility into how the model was built to fully evaluate governance, compliance and security concerns.
Read about open-weight AI models here:
How should enterprises and telcos think about open weight AI?
China races ahead in open AI — America freaks out
AT&T's AI advice: Stop 'racing from stoplight to stoplight'
OpenAI's new open weight models could change game for telcos
Model weights are key to AI’s intelligence – and an Achilles heel