Is an open-weight AI model the same as an open-source one?

In detail

Open can mean different things

A trained model is mostly its parameters, often called weights: the billions of numbers learned during training. Releasing them lets anyone download the model, run it on their own hardware and fine-tune it. Releasing them does not, by itself, tell you how the model was trained or let you rebuild it.

The Open Source Initiative’s Open Source AI Definition, version 1.0, applies the four freedoms of open-source software to AI: use the system for any purpose without asking permission, study how it works, modify it, and share it. To make those freedoms real, it requires three things. The first is data information detailed enough for a skilled person to build a substantially equivalent system. The second is the complete code used to train and run it. The third is the parameters themselves. Each has to be available under OSI-approved terms. An open-source model in that sense is a higher bar than open weights.

Three release patterns you will see

As of October 24, 2026, the model pages for three well-known families show the range:

• Open weights under a standard license. Qwen3-14B and gpt-oss-20b are published under Apache 2.0, a permissive license that allows commercial use and modification as long as you keep the license and notices.
• Open weights under custom terms. Google’s Gemma Terms of Use for Gemma 3 and 3n allow use and redistribution, but anyone who distributes one of these models or a model derived from it must include Google’s use restrictions and pass them on. The terms count offering the model as a hosted service as distribution, and they count a model distilled from these models’ outputs to reproduce their behavior as a derivative. Paige Bailey of Google DeepMind describes the Gemma family’s range: Gemma 3n at 4 billion parameters, smaller variants that fit in a browser, and Gemma 3 at 27 billion.
• Fully open. Ai2’s OLMo releases the weights together with training code, reports and the pretraining data mix.

Licenses change between model versions, so read the one attached to the exact checkpoint you plan to use.

Why teams pick open weights

Control is the usual reason. Fergal Reid describes how Intercom post-trained a 14-billion-parameter Qwen 3 model for one high-volume step in its Fin agent and gained more fine-grained control over it. He calls the open-weight release generous, given what training those models costs. The cost case is in when an open model beats a frontier API on cost.

Chainguard’s Dan Lorenc likes being able to download open models and run them on your own hardware. He also doubts the economics work for standalone companies that build open models: at billions of dollars per training run, he says, you are down to a couple of companies and governments that can afford to release the result. That is his forecast, and it is one reason to check who maintains a model before you depend on it.

What to check before you ship

Galileo CEO Vikram Chatterji says his team had to research which of the many open models it could actually use, including which had the right licenses, before building evaluation models on them. Illustrative example: a team wants to fine-tune an open model for a customer-facing assistant and sell access to it. Before training, it would check:

1. The license name and version on the exact checkpoint, not a blog post about the family.
2. Use restrictions, including any acceptable-use policy the license points to.
3. Distribution rules. Does serving the model through an API count as distributing it, and what has to travel with it?
4. Derivatives. Do the terms cover fine-tuned or distilled models, including models trained on the original’s outputs?
5. Notices and attribution you must keep.
6. Training data documentation, if an audit or a customer will ask what the model learned from.

This is a reading checklist, not legal advice. When a deal depends on the answer, involve counsel.

Where it falls short

Open weights let you inspect what a model does, not how it got that way. Without data information and training code, you cannot reproduce the training run or check the data for problems. A permissive license also says nothing about quality. Teams still need their own evaluations on their own tasks before switching. For the broader choice between open and closed models, see open vs. proprietary AI models.

Disclosure: Episodes 46, 47 and 49 are sponsored by Galileo, the Season 3 presenting sponsor. Conor announces the sponsorship at 1:08 in episode 46, and the reads are at 49:58 in episode 47 and at 28:52 in episode 49. Vikram Chatterji is Galileo’s CEO. Episode 68 is sponsored by Walrus Memory and Svix, the Season 4 presenting sponsors; Conor thanks them at 2:22.