Subscribe
,

Open Weights vs Closed Models: What “Open” Actually Means in AI

AIooze Editorial

Open Weights vs Closed Models: What "Open" Actually Means in AI (AIooze explainer illustration)

When a company says it has released an “open” AI model, it usually means one specific thing: you can download the model’s weights, the huge file of numbers that holds everything the model learned, and run it on your own hardware. That is a big deal. It is also much less than “open source” means for ordinary software, and the difference is mostly hidden in the licence.

This week alone, Meta published open weights for an 816bn-parameter model without stating a licence, Nvidia added open models to its Nemotron family, and Black Forest Labs shipped a small open-weights image model. Knowing what “open” covers, and what it does not, is now basic literacy for following AI.

The simple version

Open weights means you get the cake, not the recipe

  1. An AI model is a giant list of numbers. Training sets those numbers. They are called weights.
  2. Closed means you only get to talk to it. The company keeps the numbers and you use the model through its app or website.
  3. Open weights means you get the numbers. You can download them and run the model on your own computer.
  4. But you often don’t get the recipe. The training data and training code usually stay secret.
  5. And the rules come in a separate note. The licence says what you are allowed to do with it, like whether you can use it to make money.

Watch the narrated version2:58

More on our YouTube channel

The video loads from YouTube when you press play. Open it on YouTube

Why it matters now

Open-weights releases have become routine, and they now include very large models. Meta’s Muse Spark release has roughly 816bn total parameters, and our brief noted that its full-precision checkpoint runs to about 1.5TB. At the other end, Nvidia’s Bolt 74B uses aggressive compression to fit a 74bn-parameter model into 24GB of memory, enough for a single consumer graphics card.

Who can run a capable model, and on what terms, shapes who can build with AI, how private your data stays and how regulators respond. That is why the word “open” deserves a closer look.

How it works, step by step

Step 1: What weights are

A model such as a large language model is a network of simple calculations connected by billions of adjustable numbers, called parameters or weights. Training means nudging those numbers, over and over, until the model’s outputs get good. When training ends, the numbers are frozen and saved to a file.

That file is the model. Anyone with the file and enough computing power can run it.

Step 2: Closed models

A closed model never leaves the company that made it. You reach it through an app or an API (a way for software to send requests over the internet). The company runs the hardware, sets the prices, decides what the model will and will not do, and can change or retire it.

Think of a restaurant. You can order any dish on the menu, but you never enter the kitchen, and the menu can change tomorrow.

Step 3: Open weights

An open-weights model is one whose weights file is published, usually on a site like Hugging Face. You can run it on your own servers or laptop, inspect how it behaves, fine-tune it on your own data and keep using it even if the maker moves on.

To stretch the analogy: you get the finished cake to take home, slice and decorate however you like.

Step 4: Open source, which is a higher bar

In ordinary software, “open source” means you get the source code, the human-readable recipe, plus the freedom to use, study, change and share it. Applied strictly to AI, the recipe includes the training code and enough information about the training data to rebuild the model.

The Open Source Initiative, the body that maintains the standard definition of open-source software, published its Open Source AI Definition 1.0 on 28 October 2024. It asks for detailed information about the training data, the full code used to train and run the system, and the weights, all under terms that allow use, study, modification and sharing. Most models described as “open” publish only the weights, so by that definition they are open weights, not open source.

Step 5: The licence decides what you can do

The weights file tells you nothing about your rights. The licence does. Licences vary a lot:

  • Permissive licences, such as Apache 2.0 or MIT, let you use, change and sell products built on the model with few conditions. OpenAI released its gpt-oss models under Apache 2.0 in August 2025, for example.
  • Custom licences add conditions. Meta’s Llama 2 licence, from July 2023, required any company whose products had more than 700 million monthly active users to request a separate licence from Meta.
  • No stated licence leaves you guessing. As our Muse Spark brief pointed out, without one there is no way to tell whether the weights can be used commercially.

A worked example: a hospital chatbot

A hospital wants an assistant that summarises patient notes. Patient data cannot leave its own servers.

  • With a closed model, every note would be sent to an outside company over the internet. Even with strong contracts, the hospital’s lawyers may say no.
  • With an open-weights model, the hospital downloads the weights, runs the model on its own machines and fine-tunes it on anonymised examples of good summaries. The data never leaves the building.
  • The licence check comes first. If the licence forbids commercial or clinical use, or is missing, the hospital cannot safely proceed, however good the model is.
  • The trade-off. The hospital now runs and secures the hardware itself, and the very largest open models need far more memory than a typical server has.

What people get wrong

“Open weights means open source.” Usually it does not. Most releases share the trained file, not the training data or code, and many come with licence restrictions.

“Open means free to use for anything.” Only if the licence says so. Always read it.

“Closed models are always better.” Closed models have often led on the hardest benchmarks, but open models frequently follow within months, and for many everyday jobs a smaller open model run locally is good enough and cheaper.

“Once it is out, safety features stay on.” Anyone with the weights can fine-tune a model, including to weaken the guardrails its makers trained in. That is the central worry in debates about releasing very capable models, and one reason some labs keep their strongest models closed.

How regulators treat open weights

Governments have so far mostly chosen to watch rather than restrict.

  • United States. In July 2024 the National Telecommunications and Information Administration (NTIA) published a report on models with widely available weights. It recommended that the government not restrict the availability of open weights for currently available systems, while building the ability to monitor risks and act later if needed.
  • European Union. Under Article 53(2) of the EU AI Act, providers of general-purpose models released under a free and open-source licence, with their weights published, are exempt from some documentation duties. The exemption does not apply to models the Act classes as posing systemic risk.

Where it is heading

The direction of travel is toward more, larger and more varied open releases, including models designed specifically to run on a single machine, such as Nvidia’s Bolt. Compression formats such as NVFP4, offered alongside the full-precision weights in both the Muse Spark and FLUX.3 Klein releases, keep shrinking what it takes to run them. The open question is less technical than legal: whether licences become clearer, and whether regulators keep treating open weights more lightly as the models grow more capable.

Key terms

Sources

Last reviewed: October 2026

Related