19. Open vs Closed Weights
The open vs closed distinction affects what you can do with a model. Understanding the differences helps you choose models based on legal and technical constraints.
What "open weights" means:
Open weights models allow you to:
- Download and run the model locally
- Fine-tune on your own data
- Modify the weights
- Deploy for commercial use (varies by license)
This does not mean the training data is public, training code is shared, or the model is free to use in all contexts.
Common licensing scenarios:
| License | Models | Commercial use | Modification | redistribution |
|---|---|---|---|---|
| Apache 2.0 | Llama 3.1 | Yes | Yes | Yes |
| Llama 2 | Llama 2 | Yes (with restrictions) | Yes | No |
| CC BY-NC | Some research | No | Yes | No |
| Custom | Most closed models | No | No | No |
What "closed weights" means:
Closed models do not release weights. You can only access them through APIs or hosted services. You cannot:
- Run locally
- Inspect weights
- Fine-tune (typically)
- Deploy on your own infrastructure
The quality/efficiency tradeoff:
Closed models (GPT-4, Claude, Gemini) are generally ahead in capability because:
- Larger training compute budgets
- More sophisticated alignment
- Continuous improvement without model exposure
Open models catch up but lag by 6-12 months on average.
Running open models locally:
# Typical local inference setup
# 1. Download model
huggingface-cli download meta-llama/Meta-Llama-3.1-8B-Instruct-GGUF
# 2. Run with llama.cpp
./llama-cli -m ./llama-3.1-8b-instruct-q4_k_m.gguf \\
-n 2048 \\
-ctx 8192 \\
-p "You are a helpful assistant."
# 3. Monitor resource usage
nvidia-smi
When open weights matter:
- Privacy-sensitive data (medical, legal, financial)
- Cost optimization for high-volume inference
- Custom fine-tuning requirements
- Offline operation requirement
List 3 tasks where you must use open weights and 3 where closed models are acceptable. Document the reasoning for each.