NEW YORK — What happened
In late September 2026, the Open Source Initiative (OSI) issued definitive guidance clarifying that artificial intelligence models with commercial usage caps, user-count thresholds, or restricted deployment rights cannot be marketed as 'Open Source.' This decision directly challenges several major technology firms that release open-weight models under custom licenses that restrict large enterprise usage or commercial derivatives.
Why it matters
For years, software developers and enterprise buyers relied on the term 'open source' as a guarantee of complete freedom to inspect, modify, and run software without ongoing licensing fees or restrictions. As AI models become core infrastructure, corporations have published model weights while retaining strict control over commercial scaling. The OSI's stance establishes a clear boundary between genuine open-source AI—which requires open training code, dataset documentation, and unrestricted weight usage—and 'source-available' or 'restricted-weight' models.
Deep dive
Running models locally on company hardware gives organizations privacy and cost predictability. However, when a model is released under a license that prohibits commercial use beyond a set revenue ceiling (e.g., $10 million annual revenue) or active user count, it places legal risk on growing startups. Legal teams must now treat open-weight models with the same scrutiny as proprietary software. Developers are urged to check whether a model's weights are truly unencumbered or merely conditionally accessible for evaluation.
Report check
Claims: AI providers claim that revenue-capped licenses are necessary to stop hyper-scalers from monetizing their work without contributing back. Verification: This economic pressure is real and verified by public financial statements from model creators. However, claims that restricted licenses qualify under standard Open Source Definition guidelines have been formally rejected by international open-source standards bodies.
Open questions
Will developer ecosystems drift toward fully unencumbered open-weight models like those under Apache 2.0, or will corporate dominance force acceptance of dual-tier source-available licenses?