Open-weight model

What it means

When an AI model learns, it ends up with millions or billions of numbers called weights. The U.S. National Telecommunications and Information Administration (NTIA) describes them this way: "The weights represent numerical values that the model has learned during training."¹ Most well-known chatbots keep these numbers private. You use the model through the company's website or service, and the company sees what you send. An open-weight model works the other way: the developer publishes the weights. NTIA explains what that allows: "Open model weights allow users to perform computational inference using their own computational resources."¹ In plain words, you can run the model on your own machines.

Here is an everyday example. A hospital wants an AI assistant to summarize patient notes, but it does not want those notes leaving its building. With an open-weight model, it can download the model, run it on its own servers, and adjust it to medical language. No outside provider sees the notes.

Where the word comes from

We could not find a reliable source showing who first used the term, so we don't name anyone. Here is what we can document. In July 2023, the Open Source Initiative (OSI), which maintains the standard definition of open-source software, objected when Meta called its Llama 2 model open source. Its executive director, Stefano Maffulli, wrote: "The license for the Llama LLM is very plainly not an “Open Source” license."² By December 2023, the French company Mistral AI was describing its Mixtral release as "a high-quality sparse mixture of experts model (SMoE) with open weights."³ In July 2024, an NTIA report to the White House used the phrase "open weight models" for these systems.⁴ Today, AI labs use the term to make a narrower claim than "open source": we share the finished model, but not necessarily everything that went into making it.

How others define it

• NTIA (U.S. Department of Commerce) says weights are "widely available" when they are "open to the public by allowing users to download these weights from the Internet or through other mechanisms."⁴
• The EU AI Act gives open-source models lighter treatment, but only when they meet a test. It says such models "should be considered to ensure high levels of transparency and openness if their parameters, including the weights, the information on the model architecture, and the information on model usage are made publicly available."⁵
• The Open Source Initiative sets a higher bar for "open source AI." Weights alone are not enough. The definition also requires "Sufficiently detailed information about the data used to train the system" and "The complete source code used to train and run the system."⁶ It released this definition on October 28, 2024.⁷
• OpenAI, which mostly sells access to closed models, describes its gpt-oss models as "OpenAI's open-weight models designed for powerful reasoning, agentic tasks, and versatile developer use cases," released under the Apache 2.0 license.⁸

Where they agree and differ

Everyone agrees on the core idea: if the weights are public, you can download the model and run it yourself. The disagreements are about what else must be shared, and on what terms. The OSI's view is the strongest objection to the term: labs say "open" when they share only the finished product, while the training data and code that would let others study or rebuild the model stay private.⁶ Licenses also differ. Some, like Apache 2.0, put few limits on use.⁸ Others restrict commercial use, and the OSI rejected Llama 2's license for that reason.² Regulators look at the risks as well. NTIA warned that publishing weights "significantly lowers the barrier of entry to fine-tune models for both beneficial and harmful purposes" (fine-tuning means further training a model for a particular job). Still, NTIA did not recommend restricting these models, saying instead that "the government should actively monitor a portfolio of risks."¹⁴

In the news

On October 15, 2026, Meta released Muse Spark Open. Meta calls it "the most capable US open-weight model to date," which is the company's own claim and has not been independently tested.⁹ Meta is releasing the weights "under the Muse Model License with limited commercial terms."⁹ That detail shows why the term matters. A model can be open-weight and still come with conditions on how a business uses it. TechCrunch reported that running it locally takes "a requirement of at least 1 terabyte of RAM" (computer memory), and noted that "Open-weight models can be significantly more cost-efficient."¹⁰

What it means for your organization

An open-weight model gives you control. You can run it on your own machines, keep your data in-house, and adjust the model for your work.¹ Control also means responsibility. Once weights are published, the developer cannot recall them or monitor how they are used, so your organization has to set and enforce its own [guardrails](/definitions/guardrails) and keep [a person in the middle](/definitions/a-person-in-the-middle).¹ Read the license before you build anything on a model, because "open-weight" does not mean "free for any use."⁹

Question to ask your team: For each AI model we rely on, do we know who holds the weights, what the license lets us do, and who in our organization is responsible for how it behaves?