Open and closed models

Some AI models you can download and run on your own laptop. Others you can only reach through a company’s app or website. Both get called AI models, and both sometimes get called “open”. This lesson sorts out what each kind gives you, what you are allowed to do with it, and what you give up either way.

Closed, open weights, and open source

Closed models. A closed modelClosed modelA model whose weights are not released, so you can use it only through the company's app or API, under its terms of service, while it runs on the company's computers.Open in glossary is one you use through an app or a programming interface (an API) while it runs on the company’s computers, as in where your AI runs. You never see the weights. Terms of service decide what you may do, and the company can change the model at any time; OpenAI’s business terms, for example, say it “may update the Services periodically” (OpenAI).

Open weights. The trained parameters are published, and anyone can download them under a license. That is what open weightsOpen weightsA model whose trained parameters are published for anyone to download, under a license that says what you may do with them. It does not mean the training data or code are published too.Open in glossary means: you get the finished model, not necessarily how it was made.

Open source. The Open Source Initiative, which maintains the Open Source Definition for software, released the Open Source AI DefinitionOpen source AIAs the Open Source Initiative defines it (2024): an AI system anyone may use for any purpose, study, modify, and share, released with its parameters, its complete code, and enough information about its training data to rebuild it.Open in glossary 1.0 on October 28, 2024 (OSI). It asks for four freedoms: to use the system “for any purpose and without having to ask for permission”, to study how it works, to modify it “for any purpose, including to change its output”, and to share it. And it asks for what you would need to change it properly: the parameters, “the complete source code used to train and run the system”, and information about the training data detailed enough “so that a skilled person can build a substantially equivalent system”.

By that test, many well-known open-weights models are not open source. When the OSI tested the definition, models such as Pythia, OLMo, and T5 passed, while Llama 2, Grok, Phi-2, and Mixtral did not, because they lacked required components (OSI).

What the licenses say

Downloading the weights is only the start. A model licenseModel licenseThe terms under which a model's weights may be used, changed, and shared, from permissive open source licenses such as Apache 2.0 to licenses with use restrictions, user limits, or a ban on commercial use.Open in glossary says what you may do with them, and licenses differ more than the word “open” suggests:

• Classic open source licenses, such as MIT and Apache 2.0, allow any use, including selling, as long as you keep the copyright and license notices (Apache 2.0 also asks you to mark files you changed). They come from software, and they say nothing about a model’s outputs.
• Custom open-weights licenses add conditions. Meta’s Llama 4 license requires a separate license, at Meta’s discretion, for any company that already had more than 700 million monthly users when the model was released (one that grows that large later is not covered), and makes users follow an acceptable use policyAcceptable use policyA list of uses a model's provider forbids, such as fraud, weapons, or passing off outputs as human, which its license or terms make binding on users.Open in glossary. When Meta called Llama 2 open source, the OSI answered “it is not”, because its license restricts commercial use for some users and restricts use “for certain purposes” (OSI, July 2023). Google’s Gemma Terms of Use similarly carry a prohibited use policy, but Google released Gemma 4, in April 2026, under Apache 2.0 instead (Google).
• Responsible AI licenses, such as OpenRAIL-M, allow use and sharing like an open source license but list uses that are forbidden, even when you only run the model, and every modified version must carry the same restrictions.
• Noncommercial licenses, such as Creative Commons BY-NC, forbid use for commercial advantage. One family can mix licenses: Depth Anything V2 Small, which this site runs, is Apache 2.0, while its Base and Large versions are CC BY-NC 4.0.

Pick what you want to do and compare seven licenses and terms. Every answer quotes the clause it comes from.

# What three licenses say about four actions, read from their texts.
# "silent" means the license does not address it, so it does not forbid it.
# Llama 4's "huge" is a condition, not a no: its separate license applies only
# to companies that already had over 700 million monthly users at release.
answers = {
    "Apache 2.0": {"run": "yes", "sell": "yes", "huge": "yes", "human": "silent"},
    "Llama 4": {"run": "conditions", "sell": "conditions", "huge": "conditions", "human": "no"},
    "CC BY-NC 4.0": {"run": "conditions", "sell": "no", "huge": "conditions", "human": "silent"},
}

def verdict(license, plan):
    found = [answers[license][action] for action in plan]
    if "no" in found:
        return "not allowed"
    if "conditions" in found:
        return "allowed, with conditions"
    return "allowed"

for plan in (["run"], ["run", "sell"], ["huge"], ["human"]):
    print(" + ".join(plan))
    for name in answers:
        print(f"  {name:13} {verdict(name, plan)}")

run
  Apache 2.0    allowed
  Llama 4       allowed, with conditions
  CC BY-NC 4.0  allowed, with conditions
run + sell
  Apache 2.0    allowed
  Llama 4       allowed, with conditions
  CC BY-NC 4.0  not allowed
huge
  Apache 2.0    allowed
  Llama 4       allowed, with conditions
  CC BY-NC 4.0  allowed, with conditions
human
  Apache 2.0    allowed
  Llama 4       not allowed
  CC BY-NC 4.0  allowed

What the comparison shows:

• The classic open source licenses are short and permissive. They grant almost everything and ask you to keep notices. They were written for code and are silent about outputs.
• The newer AI licenses reach further. Llama 4’s, Gemma’s, and OpenRAIL-M’s use restrictions apply to every use, even running the model on your own computer, and OpenRAIL-M and Gemma make a service built on the model pass them on to its users. Their terms follow the model into anything derived from it, and some reach its outputs: a model trained on Llama outputs and released must have a name starting with “Llama”.
• Closed terms grant use of a service, not a model. OpenAI’s terms let you build paid products on its API and, as of January 2026, assign you its rights in the outputs, but forbid copying the service or using outputs to build competing models, with narrow exceptions.
• Several terms forbid passing off outputs as human, the provenance question from the lesson on safety in AI products. MIT and Apache 2.0 say nothing about it.

The comparison is a reading of the texts for learning, as of October 2026, not legal advice. Licenses also change between releases, as Gemma’s did.

Why this site’s models are open

Every model on this site runs inside your browser, so its weights have to be downloadable. That rules out closed models. Several labs, such as the model microscope, also read the numbers inside a model while it runs, which only works with the actual weights. Here are the site’s models and their licenses, read from its model list:

Most are MIT or Apache 2.0. SmolLM2’s authors at Hugging Face also released the datasets they built for it (Allal and colleagues, 2025). Supertonic 2, the speech model, uses OpenRAIL-M, so the text to speech lab shows its use restrictions before you download it.

The trade-offs

Neither kind is better at everything. What the evidence says, as of October 2026:

Capability. The most capable models have usually been closed, with open-weights models some months behind. Epoch AI measured that since January 2026, the best open-weights models trailed the frontier by about four months on average on its capability index (Epoch AI, May 2026), and in 2025 that open models small enough for one gaming graphics card matched the frontier after 6 to 12 months (Epoch AI, August 2025). Gaps like these depend on which benchmarks are counted.

Cost. Open weights are free to download, but you pay for the hardware and electricity to run them, and a large model needs a lot of both. Closed models charge for each use and need no hardware of your own. Which is cheaper depends on how much you use the model.

Privacy and control. A model on your own device can work without sending your data anywhere, and the version you downloaded never changes unless you change it. A closed model’s provider sees what you send and can update or retire the model. The U.S. National Telecommunications and Information Administration counted privacy among the benefits of open weights, along with letting more people and smaller organizations do research and build products, and spreading control of the market beyond a few large developers (NTIA, July 2024).

Safety. Released weights cannot be recalled, and the safety training described in the previous lesson can be removed. Lermen and colleagues (2023) undid the safety training of Llama 2’s chat models with one GPU and less than $200. Closed models are not immune: Qi and colleagues (2024) broke GPT-3.5 Turbo’s guardrails by fine-tuning it on just 10 examples through OpenAI’s own service, for less than $0.20. What matters for policy is how much extra risk open weights add compared with what people can already do. Kapoor and colleagues (2024) found that current research is not enough to measure that well, and the NTIA found the evidence insufficient to justify restricting open weights at the time, and recommended monitoring the risks instead.

Law is starting to draw these lines too. The EU AI Act exempts general-purpose models released under a free and open-source license, with their weights published, from some documentation duties, but not from the copyright policy and training data summary every provider owes, and not if the model is judged to pose systemic risk (Regulation (EU) 2024/1689, Article 53).