Open weight

Definition

A model whose trained parameters are publicly downloadable but whose training data and code may not be shared. Most 'open-source' models are technically open-weight, an important legal and strategic distinction.

Why it matters

The open-weight vs. open-source distinction matters for legal, strategic, and practical reasons. Open-weight models (Llama, Mistral) let you run and fine-tune the model, but you cannot fully reproduce it because the training data and process are proprietary. True open-source (like BLOOM or OLMo) provides everything. For enterprises, the practical difference is small: open-weight gives you data sovereignty, customization ability, and vendor independence. But for researchers, the distinction matters because you cannot audit training data for bias, verify claims about model behavior, or fully reproduce results. The labeling debate also has regulatory implications: some jurisdictions may treat open-weight and true open-source differently.

In practice

Meta's Llama models are open-weight: weights are downloadable, but training data composition is described without being fully disclosed, and the license includes usage restrictions. Mistral's models are similarly open-weight. By contrast, Allen AI's OLMo and EleutherAI's Pythia are truly open-source with full training data and code. The Open Source Initiative (OSI) released a formal definition of 'Open Source AI' in 2024 that requires training data access, putting Llama in a gray area. For most commercial users, this distinction is academic: they care about running and customizing the model, not reproducing the training run.