What happened: Mozilla Foundation released a landmark report defining what counts as "open-source AI" — drawing a hard line between truly open source (code, data, training process all open) and "open-weight" models where only the final parameters are released.
Why it matters: This definition fight determines which models qualify for regulatory exemptions, public funding, and liability shields. If the EU AI Act or US executive orders adopt Mozilla's framing, Llama 3.x and Mistral models may not count as "open source" despite their open weights.
The core distinction
- Open-source AI: Training code, data curation pipeline, pretraining data (or full provenance), evaluation methodology, and weights all under OSI-approved licenses
- Open-weight models: Final parameters released, but training data opaque, code partial or absent, reproduction impossible
Where the line falls today
Under Mozilla's rubric: zero current frontier models qualify as open-source AI. OLMo, Pythia, and a few academic projects come closest. Llama, Mistral, Qwen, Gemma — all "open-weight" only.
What we're watching
Whether OSI (Open Source Initiative) ratifies this definition. Their process is underway — if they endorse it, the term "open source AI" gets a legal meaning that changes procurement, compliance, and liability for every enterprise deploying these models.
Originally published on ayraix.com, practical AI for enterprise builders.
Top comments (0)