Definition
Model weights is Model weights are the learned numbers stored in a trained AI model that determine how it responds, and the file holding them is what a lab protects as its most valuable asset.
When an AI model is trained, it adjusts billions of numbers until its predictions get good. Those numbers are the weights. After training they are saved as large files. The code that runs a model is fairly standard. The weights are what make it this model and not another.
Weights are one kind of parameter. People often use the two words almost interchangeably, though "parameters" also covers other learned values and is usually used for the count, while "weights" usually means the actual stored numbers.
Think of a recipe versus a chef's years of practice. The recipe (the architecture and training code) can be shared in a paper. The practice, boiled down into numbers, is the weights. Whoever holds the weights can run the model.
They are not the training data, and they are not a readable database of facts. You cannot open the file and find a sentence. The knowledge is spread across the numbers in a way nobody can read directly.
If a story says a lab "protects its weights" or a company "released its weights," it is talking about the single thing that lets a model be run. Control of the weights is control of the model.
Weights are the stored numbers a trained model uses to make predictions. Parameter is the broader term, and people mostly use it for the count, such as a model with 70 billion parameters.
The trained numbers, so you can run and adjust the model yourself. It does not necessarily include the training data or training code.
Mistral opened a public preview of Mistral Large 4 on Tuesday, a mixture-of-experts model with about 1 trillion parameters and 49 billion active, and promised weights by the end of October. The API is a preview. The weights are the launch.
Reflection AI introduced Beam on Monday, a 501-billion-parameter mixture-of-experts model with 23 billion active, and promised Apache 2.0 weights later in October. Until the weights ship, it is a waitlist and a claim about efficiency.
OpenAI put its Decisions API in public beta on October 6, with gpt-6-luna at $0.10 per million input tokens and no charge for output. If you classify, route or grade at volume through a chat model, this is the endpoint to price against.
Google opened SynthID Detector worldwide, letting anyone check images, video and audio for watermarks from Google, OpenAI, Nvidia and Kakao. A yes from it means something. A no means almost nothing.
SemiAnalysis checked 857 model releases from nine Chinese developers and found a published safety result for 31 of them, and only 9 at launch. If you run Chinese open weights, the safety testing is your job.