What is an open weight model? AI you can download

What an open weight model is

An open weight model is a model whose trained weights are released to the public, so anyone can download it and run it themselves. You see the label on models such as Meta's Llama, OpenAI's gpt-oss, Alibaba's Qwen and DeepSeek. The opposite is a closed model, whose weights never leave the company that built it.

Open weight model at a glance

Also called	Open-weights model, open model
What is public	The trained weights, plus a license
What is usually private	Training code and training data
How you use it	Download it and run it, or use a hosted service
Opposite	Closed model, used only through the maker's app or API

Weights are the numbers a model learned

Weights are the values a model settles on during training, and they decide how it turns your input into an output. A large language model contains billions of them. Training is the slow, expensive part; once it is finished, the weights are saved as files.

Those files are the model in practical terms. Whoever has them can run the model. The Open Source Initiative (OSI), the group that maintains the definition of open source, notes that they can also fine-tune it, adapt it or deploy it in their own projects.

You download it instead of calling an API

With a closed model, you send your request to the maker's servers through its app or API and get an answer back. An API is the interface programs use to talk to a service. You never hold the model itself.

An open weight model can be downloaded from the internet and run on a computer you choose. The US National Telecommunications and Information Administration (NTIA) counts Meta, Google, Microsoft, Stability AI, Mistral, the Allen Institute for AI and EleutherAI among the developers that have done this, though not always with their best work. We cover who publishes the largest open models in 2026 in "Open Weight Models in 2026: Qwen and Chinese Labs."

Open weight vs open source AI

Open weight and open source are often used as if they meant the same thing, but an open weight model gives you less. You get the finished model, not what went into building it, and the table below shows where the line falls.

What each kind of model releases

	Closed	Open weight	Open source AI
Weights	No	Yes	Yes
Training code	No	Usually not	Yes
Training data details	No	Usually not	Yes
Use for any purpose	No	Depends on license	Yes

You get the weights, not the training code or data

The OSI lists what open weights leave out: the training code, the training dataset and full details of how that data was put together. Without them, nobody outside the company can rebuild the model from scratch.

That gap matters for checking a model, not for running it. If you can't see what it was trained on, you can't trace where a bias or an error came from. For everyday use the weights are enough, but for an audit or research they fall short.

Open source AI also requires the code and data details

The OSI's Open Source AI Definition 1.0 sets a higher bar than publishing weights. It requires the freedom to use the model for any purpose without asking permission, and to study, modify and share it.

To make those freedoms real, the definition asks for three things together: detailed information about the training data, the code used to train and run the model, and the weights themselves. A model that ships only weights, or whose license limits who may use it, does not meet it.

Open weights still come with a license

Being able to download a model does not mean you can use it however you like. Every open weight model ships with a license, and that license, not the download button, decides what is allowed. The two examples below sit at opposite ends.

Two open weight licenses compared

	Llama 4	gpt-oss
License	Llama 4 Community License	Apache 2.0
Commercial use	Yes, with conditions	Yes
User limit	Separate license above 700 million monthly active users	None
Attribution	Show "Built with Llama"; derived model names start with "Llama"	No naming rule

Llama needs a separate license past 700M monthly users

Meta's Llama 4 license requires any company whose products had more than 700 million monthly active users in the previous month to ask Meta for a separate license. Meta can grant it or refuse it at its discretion.

Most businesses are nowhere near that size, but the license has other conditions. If you distribute Llama or a product built on it, you must display "Built with Llama," and a model you train from it must have a name that starts with "Llama." Use also has to follow Meta's acceptable use policy.

gpt-oss uses the Apache 2.0 license

OpenAI released gpt-oss in two sizes, gpt-oss-120b and gpt-oss-20b, under Apache 2.0, a standard license that allows commercial use. There is no user cap and no rule about how you name a derived model.

Before you build anything on an open weight model, read its license on the model page. Check two points first: whether commercial use is allowed, and whether there is a cap on users or company size.

Pros and cons of open weight models

Running the model yourself gives you control over your data, but you take on the hardware and the responsibility. NTIA weighed both sides in its report on open weights and came down on watching the risks rather than restricting releases.

Trade-offs of open weight models

Pros	Cons
Data can stay on your own machines	You need hardware with enough memory
No dependence on one company's service	You run and update it yourself
Can be fine-tuned on your own data	Released weights can be misused by others

Your data stays on your own machines

Because the model runs where you put it, your prompts and files don't have to be sent to another company. NTIA counts this among the main benefits, along with spreading AI work beyond a handful of large developers.

We apply the same thinking at sakutto.ai: tools that can run AI inside your browser, such as transcription or background removal, do it on your device so your files are not uploaded. Open weights make that kind of design possible for text models too.

You need the hardware to run it

An open weight model is free to download, but it still has to run on something, and large models need a lot of memory. OpenAI says gpt-oss-20b, with 21B parameters, runs within 16GB of memory. gpt-oss-120b, with 117B parameters, needs a single 80GB GPU, a data-center card.

If you don't want to manage hardware, you can use a hosted service that runs the model for you. Coding tools can call them too, as we show with "GLM 5.2 benchmarks, pricing, and Claude Code setup." At the very large end, a model such as "Reflection Beam is a 501B open-weight MoE model" is far beyond a single machine.

Open weights can also be misused

Once weights are public, anyone can run them, including people who want to cause harm. NTIA lists misuse, gaps in oversight and unclear accountability among the risks.

After collecting 332 public comments, NTIA concluded that the evidence was not strong enough to justify restricting open weight models, and recommended that the US government monitor the risks instead. It also left open whether restrictions could be needed later.

Open weights mean you can run the model yourself

An open weight model is an AI model whose trained weights are public, so you can download it and run it yourself instead of going through the maker's app or API. It is not open source AI: the training code and data usually stay private, and the license still decides what you can do.

Before you use one, check the license for commercial use and user limits. The payoff is control over your data; the price is the hardware to run it.