Open weights versus closed models: what gets released and why
Open-weight models release the trained parameters (weights) publicly, allowing anyone to download, modify, and run the model. Closed models keep weights private and provide access only through an API or limited interface. The choice depends on trade-offs between transparency, control, safety, and commercial interests.
Defining open-weight and closed models
An open-weight model is one whose trained parameters, called weights, are made publicly available for download. The NTIA's open model weights report defines these as dual-use foundation models with widely available model weights, noting that developers such as Meta, Google, and Mistral have released models this way, though not always their most advanced ones. Open-weight models allow users to run the model on their own hardware, fine-tune it, and inspect its behavior.
A closed model, by contrast, keeps its weights proprietary. Access is typically provided through a paid API, a web interface, or a limited trial. The inner workings, training data, and architecture are not disclosed. Examples include many commercial offerings where the vendor retains full control over the model's use and distribution.
The term 'open source' is sometimes used for open-weight models, but it is not identical. As IBM's article on open source LLMs explains, open source LLMs refer to models whose code and underlying architecture are publicly accessible, allowing anyone to use, modify, and distribute them. However, the license may impose restrictions, and not all open-weight models meet the official Open Source Definition.
What gets released in open-weight models
When a developer releases an open-weight model, they typically publish the weight files, a model card, and a license. The Hugging Face Hub documentation states that models are stored in repositories with attributes that make exploring and using them easy. The model card, as described in the Hugging Face model cards documentation, is a Markdown file that describes the model, its intended uses, limitations, training parameters, datasets, and evaluation results.
The weight files themselves are large binary files containing the numerical parameters learned during training. For example, the Llama 2 paper on arXiv describes releasing models ranging from 7 billion to 70 billion parameters. Alongside the weights, developers often share a configuration file, tokenizer, and sometimes fine-tuned versions optimized for specific tasks like dialogue.
A license defines what users can do with the model. Some open-weight models use permissive licenses like Apache 2.0, while others have custom licenses that restrict commercial use or require attribution. The Hugging Face Hub displays the license in the model card metadata, making it easy for users to check before downloading.
Why developers choose open or closed release
Developers release open-weight models for several reasons. The NTIA report highlights that open models diversify participation in AI research, decentralize market control, and enable users to keep data confidential by running models locally. Open release can also accelerate innovation by allowing the community to build on the work, as seen with the Llama 2 release, which the authors hoped would enable responsible development of LLMs.
Closed models are chosen when the developer wants to protect intellectual property, maintain control over use, or monetize access through subscriptions or usage fees. Keeping weights private reduces the risk of misuse or unauthorized modification. The NTIA report notes that many developers provide limited or no public access to their advanced models, including their weights, to manage safety and accountability.
Some organizations adopt a hybrid approach, releasing an older or smaller model openly while keeping the most capable version closed. This allows them to gather community feedback and build trust without exposing their highest-performing system. The choice is influenced by the balance between promoting innovation and mitigating potential harms.
Benefits and risks of open-weight models
Open-weight models offer significant benefits. They allow researchers with limited resources to experiment and contribute, as noted by the NTIA. They also enable transparency: users can inspect the model's behavior, audit for biases, and verify claims. IBM points out that open source LLMs provide flexibility, cost savings, and the ability to fine-tune on proprietary data without sharing it with a third party.
However, open weights also introduce risks. The NTIA report warns of potential harms to national security, equity, safety, privacy, and civil rights through misuse, lack of oversight, or unclear accountability. Because anyone can download and modify the model, malicious actors could use it for harmful purposes without detection. The report concludes that current evidence does not warrant blanket restrictions but recommends ongoing monitoring and preparedness.
Closed models mitigate these risks by limiting access, but they concentrate power in a few companies and reduce transparency. Users must trust the vendor's safety measures and cannot independently verify claims. The debate between open and closed release is ongoing, with policymakers, researchers, and industry stakeholders weighing the trade-offs.
| Aspect | Open-weight model | Closed model |
|---|---|---|
| Access to weights | Publicly downloadable | Not released; access via API |
| Modification | Allowed (fine-tuning, adaptation) | Not possible |
| Transparency | High: model card, architecture, training data often disclosed | Low: internal details proprietary |
| Cost | No licensing fee; infrastructure cost for running | Pay-per-use or subscription |
| Control | User controls deployment and data | Vendor controls access and usage |
| Safety oversight | Community and self-governance | Vendor manages safety filters |
| Examples | Llama 2, Falcon, Mistral | GPT-4, Claude, Gemini |
Frequently asked questions
What is the difference between open-weight and open-source models?
Open-weight means the trained parameters are publicly available, but the license may not meet the official Open Source Definition. Open-source models typically require that the code, architecture, and training data are also accessible and that the license allows free use, modification, and distribution.
Can an open-weight model be used for commercial products?
It depends on the license. Some open-weight models use permissive licenses like Apache 2.0 that allow commercial use, while others have restrictions. Always check the model card and license before using a model commercially.
Why would a company release a model with open weights?
Companies release open-weight models to foster innovation, gain community feedback, build trust, and decentralize AI control. It can also help smaller organizations and researchers participate in AI development without relying on a single vendor.
What are the main risks of open-weight models?
Open weights can be misused for harmful purposes, such as generating disinformation or malicious code. They also make it harder to enforce safety measures and accountability, as anyone can modify and redistribute the model.
How can one find and download an open-weight model?
Platforms like the Hugging Face Hub host thousands of models. Each model has a repository with a model card that describes its capabilities, license, and usage instructions. One can filter by task, license, or framework to find a suitable model.