What are Open Weights vs Open Source?
In the AI world, open weights means a model's trained parameters are publicly available for download and use, while open source traditionally means the complete package — code, data, training methodology, and weights — is available under a license that allows modification and redistribution. Most models marketed as "open source" are actually open weights only.
Why It Matters
The distinction matters for trust, reproducibility, and rights. True open source lets anyone reproduce, modify, and understand the model fully. Open weights only lets you use and fine-tune the model — you can't verify training data, reproduce training, or fully audit for biases. The terminology debate also has legal implications: the Open Source Initiative argues that "open source AI" requires more than just releasing weights.
How It Works
Open weights (most "open" models):
- Model weights (parameters) are downloadable
- You can run inference locally
- You can fine-tune for your use case
- Training data is NOT provided
- Training code may or may not be provided
- Often has usage restrictions (license limitations)
- Examples: LLaMA 3 (Meta), Gemma (Google), Mistral, Command R
True open source:
- Model weights available
- Training code available
- Training data available (or described in detail)
- Permissive license (Apache 2.0, MIT)
- Anyone can reproduce the full training run
- Examples: OLMo (AI2), Pythia (EleutherAI), BLOOM (BigScience)
Proprietary (closed):
- Only accessible via API
- No weights, no training details
- Examples: GPT-4 (OpenAI), Claude (Anthropic), Gemini (Google)
The spectrum of openness:
- Fully closed (API only)
- Weights available, restrictive license
- Weights available, permissive license
- Weights + training code
- Weights + code + data + permissive license (true open source)
Why it matters in practice:
- Open weights enable: local deployment, fine-tuning, privacy compliance
- True open source enables: auditing, reproducibility, trust, community improvement
- License restrictions affect: commercial use, derivative works, liability
Example
Meta's LLaMA 3 is open weights: you can download the model, run it locally, and fine-tune it, but under a license that restricts use above 700M monthly users and doesn't provide training data. AI2's OLMo is closer to true open source: weights, training code, training data, and evaluation code are all publicly available under Apache 2.0.