The categories that matter
Three categories matter and are often confused in casual writing.
| Type | Weights | Training code | Data | Modify & redistribute? |
|---|---|---|---|---|
| Open Source (OSI definition) | Available | Available | Available or specified | Yes, fully |
| Open Weight | Available | Usually not | Usually not | Varies — check license |
| Closed / Proprietary | Not available | Not available | Not available | API access only |
Most "open" LLMs are open-weight
LLaMA, Mistral, Qwen, DeepSeek — all release weights under custom community licenses with restrictions (commercial caps, naming requirements, compliance limits, sometimes sublicensing rules). They are not OSI-open-source. They are open-weight under specific terms.
True open-source LLMs are rare
OLMo from AI2 is the cleanest example: weights, training code, and full training-data documentation under permissive licenses. Smaller research-lab releases (BLOOM, Pythia) also qualify. Most production-grade LLMs do not.
Why this matters for production
- Commercial use restrictions. Llama community license has clauses around 700M MAU. Some research-only licenses prohibit any commercial use.
- Distillation restrictions. Some licenses prohibit using the model's outputs to train other models. This affects synthetic-data pipelines.
- Naming and attribution. Llama-derived models often must include "Llama" in the name. This sounds minor but matters for branding.
- Sublicense terms. If you distribute a fine-tune, what terms apply downstream? Often the original license must propagate.
The reading rule
Read the actual license file before commercial deployment. "Open-source" claims in marketing are often inaccurate. The license itself is the contract; the marketing language is not.