AI models broadly come in two flavors. Closed models, such as those behind ChatGPT, Claude and Gemini, are accessed only through their makers’ apps and APIs. Open-weight models, such as Meta’s Llama family, Mistral’s open models, Alibaba’s Qwen, DeepSeek’s models, Google’s Gemma and OpenAI’s gpt-oss models, released under the permissive Apache 2.0 license in August 2025, can be downloaded and run on your own hardware.
Updated September 2026: we added links to the Open Source Initiative’s definition, example model licenses and OpenAI’s gpt-oss release.
What “open weights” means
A model’s weights are the billions of numbers learned during training; they are the model. Releasing the weights lets anyone run and modify the model. That is different from fully open-source AI. The Open Source Initiative’s Open Source AI Definition, published in October 2024, also requires the complete training code and “sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system.” Most open-weight models do not release their full training data, and licenses vary. Meta’s Llama 3.1 license, for example, requires companies with more than 700 million monthly active users to ask Meta for permission, while gpt-oss uses the unrestricted Apache 2.0 license. Always read the license.
The trade-offs
| Factor | Closed models | Open-weight models |
|---|---|---|
| Top-end capability | Usually the frontier | Often close behind, sometimes competitive |
| Setup | Sign up and use | Requires hardware or a hosting provider |
| Data control | Data goes to the provider | Can stay entirely in your environment |
| Customization | Limited to provider options | Full fine-tuning possible |
| Cost | Pay per use or subscription | Pay for compute; can be cheaper at scale |
| Safety features | Built in and maintained by provider | Your responsibility |
| Updates | Automatic | You manage versions |
When open-weight models make sense
- Strict data requirements, such as healthcare, defense or finance, where data must not leave your infrastructure
- High volume, narrow tasks where a smaller, fine-tuned model is cheaper than a frontier API
- Offline or edge deployment on devices without reliable internet
- Research and experimentation, where access to the model’s internals matters
- Avoiding lock-in to a single vendor
When closed models make sense
- You need the most capable model available for complex reasoning, writing or coding
- You do not have the expertise or budget to run infrastructure
- You want built-in features such as search, file handling and agents
- Speed to launch matters more than control
Running models yourself
Smaller open-weight models can run on a capable laptop using free tools such as Ollama or LM Studio; OpenAI says gpt-oss-20b runs on devices with 16 GB of memory. This is a good way to experiment privately, as we discuss in our on-device AI comparison. Larger models need powerful GPUs, which is why many businesses use cloud providers that host open models for them.
The debate Supporters argue that open weights spread innovation, enable independent research and reduce concentration of power. Critics worry that openly released capable models cannot be recalled if misused. Both sides raise serious points, and policymakers are still working through them.
Choosing between them
Many organizations end up using both: closed frontier models for the hardest general tasks and open-weight models where cost, privacy or customization matter most. The right choice depends less on ideology than on your data, budget and technical capacity.
Sources
- The Open Source AI Definition 1.0, Open Source Initiative, October 2024
- Introducing gpt-oss, OpenAI, August 2025
- Llama 3.1 model card and license, Meta via Hugging Face
