Foundations
Open Weight vs Closed Weight Models
What it actually means when a model is "open source" — and why most flagship models aren't.
What "weights" are
A trained model's weights are the (typically billions of) numerical parameters learned during training — the actual thing that makes the model work. Downloading a model's weights means you have a real copy of the model itself, not just access to it through someone else's API.
Open weight
An open weight model publishes those parameters for anyone to download, run on their own hardware (or a cloud provider of their choice), and in many cases fine-tune further. Meta's Llama family, Mistral's smaller models, Alibaba's Qwen family, and DeepSeek's models are commonly cited open-weight examples. Note that "open weight" isn't quite the same claim as "open source" in the traditional software sense — the training code and, especially, the training data are usually still kept private even when the weights are freely downloadable, and the license attached to the weights can still restrict certain commercial uses.
Closed weight
A closed weight model is only accessible through the provider's own API or products — you send requests to their servers and get responses back, but never obtain the parameters themselves. GPT-4o, Claude, and Gemini are all closed weight: OpenAI, Anthropic, and Google each control access, pricing, and availability directly, and reserve the right to change or retire a model version at any time.
The practical trade-offs
- Control and privacy — open weights can be run entirely on infrastructure you control, which matters for organizations that can't send data to a third party's servers.
- Customization — open weights can be fine-tuned on your own data far more directly than closed models typically allow.
- Frontier capability — the very largest, highest-scoring models on most benchmarks have usually been closed weight, though the gap between the best open and closed models has narrowed significantly over time.
- Cost and ops burden — running a large open-weight model yourself requires real GPU infrastructure and expertise; a closed model's API trades that complexity for a per-token bill.