<p>When AI companies say their models are "open source," they increasingly mean something narrower than the software industry's traditional definition. The fracturing of what "open" means has practical consequences for organizations building on these models — particularly around compliance, fine-tuning rights, and competitive differentiation.</p>
<h2>The Spectrum of Openness</h2>
<p>At the most closed end: proprietary APIs (GPT-4o, Claude, Gemini) — you get inference access, nothing else. Moving toward open: open weights with restricted commercial use (Llama 2 original, some Mistral variants) — you can run the model, modify it, but can't compete with Meta using it. Further open: open weights with permissive commercial use (Llama 3, Llama 4) — you can build commercial products, but you don't have the training data or recipe to reproduce the model from scratch. Fully open: weights + training data + training code (OLMo from AllenAI, Falcon from TII) — everything published, model fully reproducible.</p>
<h2>Why It Matters</h2>
<p>The distinction matters for several reasons. Fine-tuning rights: open weights let you fine-tune, but without training data, you can't know what biases or knowledge gaps you're starting from. Compliance: regulated industries (banking, healthcare, legal) need to document what their AI systems were trained on — "the training data is proprietary" is an acceptable answer internally; it's problematic for external audit. Competitive moat: building a product on Llama creates a dependency on Meta's goodwill and future licensing decisions.</p>
<h2>The Trend</h2>
<p>True open source AI — weights, data, code — is growing but slowly. The models that combine commercial competitiveness with full openness don't yet exist at frontier scale; the training data cost alone ($100M-1B for frontier models) makes open data releases economically difficult. The Open Source Initiative has published criteria for what counts as truly open AI, and most "open" models don't meet them.</p>