Large language models once seemed largely the province of major technology companies and well-funded startups because training them requires substantial data and computing power. Releases such as Meta’s Llama series and DeepSeek’s V3 and R1 models have helped broaden participation. Open source software makes its source code publicly available for reuse and modification. The open source movement also holds that sharing can lower barriers to entry and improve software through collaboration. section 4.12
For AI models, “open” needs qualification. Strictly speaking, an open source model has a license allowing full use, modification, and redistribution. An open-weight model shares its parameters—the values shaped during training—but may keep its training data, methods, or pipelines closed. Some open releases have also relied on outputs from proprietary systems, so open and closed development are not always separate. section 11.4
Openness brings practical choices as well as benefits:
- Shared models can support experimentation, customization, and wider access. Stable Diffusion’s 2022 release, for example, was followed quickly by new interfaces, fine-tuned versions, and content filters. section 11.4
- A downloadable model can require local computing resources, setup, and deployment. Safety practices vary, and licenses may be vague or restrictive; wider access can also increase the risk of misuse. 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 large language models across transparency, access, safeguards, licensing, customization, infrastructure, and misuse risk. The comparison helps make clear why neither approach is an uncomplicated solution. section 11.4
Open source communities offer a path toward transparency, collaboration, and local control, but openness alone does not settle questions of safety or equitable access. Computing resources remain unevenly distributed, shaping who can build and benefit from AI. section 11.6 section 11.7
