NEW YORK — What happened
Meta has released its long-anticipated Llama 4 model family, retaining an 'open-weight' classification rather than true open source. However, the accompanying license agreement introduces a new revenue threshold clause that requires cloud giants to negotiate distinct commercial terms if they bundle the model into managed infrastructure services.
Why it matters
For smaller startups and independent software developers, the weights remain free to download, fine-tune, and self-host. But the shift targets infrastructure aggregators who profit off raw model hosting without contributing back to core research or safety frameworks.
Deep dive
The debate centers around the OSI definition of open source, which requires modification and redistribution rights without discrimination against fields of endeavor. Because Llama 4 restricts deployment based on monthly active user counts and cloud service packaging, software freedom advocates argue it is source-available proprietary software, while corporate defenders call it a pragmatic compromise that keeps foundation models accessible.
Report check
Claims: Meta is entirely locking down its AI models behind paywalls. What is verified: The base weights are downloadable on Hugging Face, but commercial usage above 700 million monthly active users now triggers mandatory licensing talks. Rumor: Competitors are preparing a legal challenge regarding antitrust concerns over cloud bundling.
Open questions
How will regulatory bodies in the EU and US classify these hybrid licenses under upcoming artificial intelligence acts, and will truly open-weights models catch up in raw benchmark performance?