Could you download an AI model, adapt it, and share your version? An open-source model is released under a license permitting use, modification, and redistribution. An open-weight model is different: its learned parameters are shared, but its training data, methods, or pipeline may remain closed. A release described as “open” therefore does not necessarily reveal how the model was made. section 11.4
Access can invite rapid experimentation. After Stable Diffusion was open-sourced in 2022, people produced new interfaces, customized versions, content filters, and other variants. Open language models have likewise grown through community development. Some now rival—and sometimes surpass—proprietary models, though some notable open releases have relied on outputs from closed systems during training or evaluation. Open and closed AI progress are not entirely separate. section 11.4
Table 11.2 Comparing open and closed LLMs
| Trade-offs |
Open source LLMs |
Closed source LLMs |
| Transparency |
Full access to weights, architecture, and sometimes data |
Opaque methods, weights, and training data |
| Accessibility |
Downloadable, runnable offline, and modifiable |
API access controlled by the provider |
| Safety alignment |
Varies widely; depends on developer practices and community norms |
Extensive internal testing and standardized safeguards |
| License clarity |
Apache, MIT, RAIL, custom terms; sometimes vague or restrictive |
Proprietary terms of service |
| Customization |
Fully customizable via fine-tuning and architectural changes |
Limited to API parameters or provider tooling |
| Infrastructure burden |
Requires local compute, setup, and deployment |
Hosted by the provider |
| Risk of misuse |
Higher; fewer enforced safeguards, and fewer default protections |
Moderated through platform enforcement and terms of service |
Compare open and closed language models across access, customization, safety practices, licensing, and the computing resources needed to run them. section 11.4
The trade-offs extend beyond performance. Downloadable models can be modified or run offline, but users must provide computing resources and handle setup and deployment. Safety protections vary, and fewer enforced safeguards can increase misuse risks. Open releases can also accelerate innovation while making intellectual-property rights and control over a model’s use harder to enforce. Efforts to make large language models cheaper and more accessible are promising, but their success remains uncertain. section 11.4 section 1.14 section 4.15
