What is an open-weight model, and is it really open source?

An open-weight model is one whose trained parameters you can download and run yourself. That is not the same as open source: most open-weight models come without their training data, and many come with licences that limit how you may use them.

What are a model's weights?

Weights, also called parameters, are the very large set of numbers that make up a language model. Training adjusts those numbers until the model predicts text well. Once training ends, the weights are the model. With the weight files and the code to run them, you can generate text on your own hardware.

Size is counted in parameters. OpenAI's gpt-oss-120b, for example, has 117 billion parameters, of which 5.1 billion are active for each token, and OpenAI says it fits on a single 80GB GPU.

How do closed, open-weight and open source models differ?

The difference is what you get: a closed model stays on the provider's servers, open weights can be downloaded and run yourself, and open source adds enough about the training data to rebuild a similar system.

With a closed model you send text to the company's servers and get an answer back. You never hold the model, and the provider can change it, reprice it or retire it. With open weights, a copy sits on your disk and keeps working the same way for as long as you keep it.

What counts as open source AI?

Under the Open Source Initiative (OSI) definition, an open source model must come with enough design information that someone could substantially recreate it, details of its training data, and freedom to use and modify it for any purpose. The OSI released version 1.0 of its Open Source AI Definition on 28 October 2024, according to TechCrunch. The training data details must say where the data came from, how it was processed and how it can be obtained or licensed. Users must not have to ask permission. The initiative has no power to enforce this.

Meta's Llama models fall short, TechCrunch reported, partly because the licence requires companies with more than 700 million monthly active users to ask Meta for a separate licence. Meta, which took part in drafting the definition, said it disagreed with it.

What do popular open-weight models actually release?

It varies a lot: only one of the five models below publishes weights, training code and named training datasets. What follows comes from each model's card on Hugging Face.

Apache 2.0 and MIT are standard permissive software licences. They allow commercial use, changes and redistribution with few conditions. DeepSeek's card says R1 may be modified and used to train other models through distillation, where one model learns from another's outputs.

Custom licences add rules. The Llama 3.1 licence requires products to display "Built with Llama", requires any model trained on Llama outputs and then shared to start its name with "Llama", sets the 700 million user threshold, and brings in an acceptable use policy. Downloading the weights also means sharing your contact details with Meta.

NVIDIA's licence allows commercial use and says NVIDIA claims no ownership of outputs. But your rights end automatically if you bypass or weaken the model's safety guardrails without a similar replacement, or if you sue claiming the model infringes a patent or copyright. NVIDIA can also update the terms, and you must accept them or stop using the model.

Only OLMo 2 publishes the full set: weights, training code and named training datasets. Ai2 releases its code and model under Apache 2.0.

Why do labs give their weights away?

Meta's chief executive Mark Zuckerberg argued in July 2024 that open models give companies more control, cost Meta no revenue and are safer. Companies can fine-tune an open model on their own data without the model's maker seeing it, and avoid depending on a vendor that might change its terms. Meta, he wrote, does not sell access to its models, so releasing them does not cut into its revenue. He also argued that models open to wide inspection are safer, and that keeping them closed would not stop rival states from getting them.

Our take: releasing weights also earns a lab developers, research attention and a say in which tools become the standard. Releasing training data is a different matter. It shows exactly what went in, which is where the competitive and legal risks sit, and that explains why so few labs do it.

How should you read an "open" model launch?

Start with the licence, then check what was actually released.

1. Find the licence name on the model card. Apache 2.0 or MIT means few strings. A company-named licence means read it in full.
2. Look for use limits. User thresholds, naming rules, acceptable use policies and termination clauses all count.
3. Check what was actually released. Weights alone let you run and fine-tune a model. Training data and code let you study it and rebuild it.
4. Treat "open source" as a claim to check, not a description, unless the training data is published too.