Open-weight model: what it means, explained simply

In plain words

An AI model is, in the end, a huge file of numbers called weights or parameters. They are what the model learned during training. Whoever has the file can run the model.

Some companies keep that file to themselves. You can use the model only through their app or API. Others publish it for anyone to download. That is an open-weight model: you can run it on your own GPU, keep your data private, and fine-tune it for your own task. Llama, DeepSeek and OpenAI's gpt-oss are well-known examples.

Open weights do not always mean everything is open. Many releases keep the training data and training code private, and some licences limit who may use the model or for what. That is why careful writers say open-weight rather than open source.

Everyday example

A sweet shop shares its laddoos, not its recipe

A famous sweet shop lets anyone take home trays of its laddoos, resell them, even add their own flavour on top. That is generous. But the recipe, the suppliers and the kitchen notes stay secret, so nobody can make the same laddoo from scratch. An open-weight model is the tray of laddoos. Open source, in the strict sense, would also hand over the recipe.

Where the example breaks: Laddoos get eaten. Downloaded weights can be copied endlessly and changed in ways the maker cannot undo or track, which is both the point of openness and the main safety worry about it.

Go deeper

How it actually works

A trained model is made of three things: the weights, the code that trains and runs it, and the data it learned from. Openness is about which of these you receive, and what the licence lets you do.

Closed models
The weights stay on the company's servers. You send requests and get answers back. You cannot inspect or change the model, but the company can update, monitor and restrict its use.
Open-weight models
The weights are published, usually with code to run them. You can host the model yourself, run it offline, compress it for smaller machines or fine-tune it. Training data and full training code are usually not released.
The licence decides the rest
Licences range from fully permissive (Apache 2.0 for gpt-oss, MIT for DeepSeek-R1) to custom ones with conditions. The OSI has said Meta's Llama licence is not open source because it restricts some commercial use and fields of use.
Fully open models
A few projects release everything: weights, training code and the training data itself. The Allen Institute for AI's OLMo models are a well-known example, built so researchers can study every step.

Open Source AI Definition: Version 1.0 published by the Open Source Initiative on 28 October 2024
What OSI requires: Freedom to use, study, modify and share for any purpose, plus the weights, the full code, and detailed information about the training data
OpenAI's return to open weights: gpt-oss-120b and gpt-oss-20b, August 2025, under Apache 2.0: its first open-weight language models since GPT-2
DeepSeek-R1: Weights released under the MIT licence, which allows commercial use and modification

Is an open-weight model open source?

One view

Close enough to call it open source

Some model makers, notably Meta, describe their open-weight models as open source AI. Meta says open source will ensure more people around the world have access to the benefits of AI. The argument: anyone can download, run, study and build on the weights for free, which delivers most of what people want from openness.

The other view

No: open source needs more

The Open Source Initiative, which maintains the long-standing definition of open source software, says open source AI must also provide the training code and enough detail about the data to rebuild a similar system, with no limits on who uses it or for what. By that test, many open-weight releases fall short.

Most likely: Both sides agree that downloadable weights are a big step towards openness. The precise term for weights without data, or with use restrictions, is open-weight; open source AI, in the OSI's sense, means the whole recipe. Everyday usage remains loose, so check the licence and what was actually released.

Questions people ask

What is the difference between open-weight and open source AI?

Open-weight means the trained model file is published so you can run and adapt it. Open source AI, as defined by the Open Source Initiative in 2024, also requires the training code, detailed information about the training data, and a licence with no limits on use. Every open source model has open weights, but not the other way round.

Is Llama open source?

Meta calls Llama open source and publishes its weights. The Open Source Initiative disagrees, because Llama's licence restricts some commercial use and certain uses, and the training data is not released. Most precise writers call it an open-weight model.

Can you run an open-weight model on your own computer?

Often, yes. Small and compressed models run on a good laptop or even a phone. Large ones need powerful GPUs with lots of memory. Once downloaded, the model runs offline and your prompts never leave the machine.

Are open-weight models safe?

It is debated. Openness lets outside researchers inspect models and lets people keep their data private. But once weights are public, anyone can fine-tune away the guardrails, and the maker cannot recall or patch the copies already out there.

Why do companies give away their model weights?

Makers give different reasons. Meta, for example, argues openness spreads the benefits of AI more widely and helps developers build on its models. Open weights also let businesses run a model on their own hardware and keep their data in-house, which can win users from closed rivals.

Open-weight model