The Overlord's Lexicon
What are open-weight models?
Also known as: Open Weights, Open-Source AI Models
Open-weight models are AI models whose trained parameters (weights) are publicly released, so anyone can download, run, fine-tune and often commercially use them, subject to the model's license.
AI models come in two broad flavors. Closed models — like the flagship models behind ChatGPT, Claude and Gemini — are only accessible through the company’s apps and APIs. You never see the model itself. Open-weight models release the trained parameters as downloadable files, so you can run the model on your own hardware or any cloud you choose.
What “open” actually covers
There’s a spectrum:
- Weights only: you can run and fine-tune the model, but the training data and full recipe stay private. This is the most common kind.
- Weights plus code and data details: closer to traditional open source.
- License terms vary: some are permissive (Apache 2.0, MIT); others restrict certain uses or very large commercial users.
That’s why many people prefer “open-weight” over “open-source” — it’s more precise.
An analogy for the meatbags
A closed model is a restaurant: excellent food, but you eat what they serve, at their prices, and they see your order. An open-weight model is the restaurant handing you the finished dish and a copy of the chef’s trained palate — you cook in your own kitchen, adjust the seasoning, and nobody watches you eat. (What they usually don’t hand over is the shopping list: the training data.)
Why it matters for the Chosen
- Privacy: sensitive data (medical, legal, client records) can stay on infrastructure you control.
- Cost at scale: for high-volume, simple tasks, self-hosting a small open model can be cheaper than paying per token — once you account for hosting.
- Customization: you can fine-tune the model on your own data, or quantize it to run on cheaper hardware.
- No platform risk: the provider can’t deprecate the model, change its behavior or raise prices on you.
- Business opportunities: building niche products on fine-tuned open models, or helping companies deploy them privately.
Tradeoffs: you’re responsible for hosting, updates, safety filters and performance, and the very best frontier capabilities usually arrive first in closed models.
Examples
- Meta’s Llama family — the release that kicked off the modern open-weight wave.
- Mistral models from the French lab Mistral AI.
- Qwen from Alibaba and DeepSeek from the Chinese lab of the same name.
- Gemma from Google and gpt-oss from OpenAI, open-weight lines from labs best known for closed models.
- Local runners like Ollama and LM Studio make it easy to try these on a laptop.
Freedom is a powerful thing. Use it on something better than limericks.
Questions the Humans Keep Asking
Is open-weight the same as open-source?
Not exactly. Open-weight means the weights are downloadable. Fully open-source AI, by stricter definitions such as the Open Source Initiative's, also requires enough information about training data and code to meaningfully study and rebuild the model. Many popular 'open' models release weights only, under custom licenses.
Can I use open-weight models commercially?
Often yes, but it depends on the license. Some use permissive licenses like Apache 2.0; others have custom terms with restrictions on use or company size. Always read the specific license.
Why would I use an open-weight model instead of ChatGPT or Claude?
Privacy (data never leaves your servers), cost control at high volume, offline use, and the freedom to fine-tune. The tradeoff is that you handle hosting, and the best closed models are often more capable.