MMOOZZGG.com

Glossary

What is an open weights model?

In short

An open weights model is a neural network whose parameter files you can download and run on your own hardware.

Language models · open-weights model

MMOOZZGG.com editorsChecked: 1 min read

For a child

Some games you can play only at a friend’s house. And some your friend gives you a copy of: play at home as much as you like and even make up your own rules. Neural networks are the same. Some you can take home and use without asking for permission every time.

In plain words

The skills of a neural network are stored in its weights: a huge set of numbers that training produces. With closed models, the weights stay with the developer, and you can use them only through its website or API. An open weights model publishes this file: people download it and run it on their own computer or server, even without the internet, and fine-tune it for their task. The training data and code often stay closed. It is like a finished cake that you get whole, but without the recipe and the list of ingredients.

For business

Open weights give a company control: the model runs on its servers, customer data goes nowhere, and the price does not depend on the rates of an outside service. This is a direct path to meeting personal data localization rules and protection against a vendor that cuts access or changes its terms. In return, the company pays for hardware, setup and support, and on complex tasks these models can be weaker than the leading closed ones. The main thing before a launch: read the license. Some models are released under free licenses, others limit commercial use, the number of users or the fields of use. How to run such a model in-house and what it costs is covered in the section on your own AI.

For a specialist

An open-weights model publishes its trained parameters (weights) for download, usually with the architecture, the tokenizer and the inference code. This is not the same as open source in the strict sense. The Open Source AI Definition 1.0 from the Open Source Initiative requires the freedom to use, study, change and share the system, and access to three components: the parameters, the full training and inference code, and enough information about the training data. Many popular models publish only the weights under their own license with limits. Example: Llama 2 (2023), a family from 7 to 70 billion parameters; its license requires a separate permission from Meta for services with more than 700 million monthly users. Weights are shared in the safetensors and GGUF formats, often in quantized versions; fine-tuning uses LoRA and QLoRA. Before use, people study the model card and the license terms.

Example

A clinic wants AI to sort medical records but cannot send them to an outside service. It downloads an open weights model, runs it on its own server and fine-tunes it on anonymized examples.

Related terms

Questions and answers

How is an open weights model different from open source? Open weights mean that the finished model file is published. Open source in the strict sense also requires the training code, information about the data and a free license.

Many open models give only the weights.

Can you use an open model for commercial purposes? It depends on the license of the specific model. Some models allow any use, others have limits on the number of users or on fields of use, or require you to credit the author.

Read the license before the launch.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us