Open License
An Open License is a legal instrument that grants users specific freedoms to use, modify, and distribute creative works or software, typically without royalty payments. In the context of artificial intelligence, it defines the terms under which models, datasets, and weights can be accessed and deployed.
Key Characteristics
- Freedom to Use: Permission to run the code/model for any purpose.
- Freedom to Study/Modify: Access to source code or weights to understand and alter functionality.
- Freedom to Redistribute: Right to share original or modified versions with others.
- Non-Discrimination: Must not restrict fields of endeavor or specific individuals/groups.
Distinctions
- Open Source vs. Open License: Open Source Software adheres strictly to OSI-approved licenses, whereas “open weights” in AI may use permissive but non-OSI-compliant licenses (e.g., restricted commercial use).
- Copyleft vs. Permissive: Copyleft licenses require derivatives to remain under the same license; Permissive Licenses (e.g., MIT, Apache 2.0) allow relicensing into proprietary forms.
Recent Developments in AI Licensing
The definition of “open” in large language models is evolving due to computational requirements and safety concerns.
- NVIDIA Nemotron 3 Ultra Case:
- NVIDIA released the Nemotron 3 Ultra model under an open license framework, positioning it as a free resource for developers.
- An independent assessment highlighted the tension between openness and practical limitations.
- Key findings from NVIDIA Nemotron 3 Ultra: Independent Assessment of Capabilities, Coding Flaws, and Open License indicate:
- The model is marketed as “free and open,” yet assessments reveal specific coding flaws.
- There is a discrepancy between the marketing narrative (“Gift To All Of Us”) and the technical reality of deployment constraints.
- The license structure allows broad access but may contain implicit restrictions or limitations on commercial derivative works, challenging the traditional open-source definition.
Common AI Licenses
- Apache 2.0: Permissive, includes patent grant.
- MIT: Minimal restrictions, permissive.
- Llama Community License: Open for use but restricts large-scale commercial deployment without permission (not strictly open source).
- RAIL (Responsible AI License): Adds usage policies to standard licenses, restricting harmful applications while keeping weights accessible.