What does “open” mean for an AI model?

2 Sents

A weights download, its license and its training materials answer different questions. I want to check each before choosing a model and deciding where to run it.

This retrospective entry follows an October 16 request. It was written and first published October 27, 2026, when the linked sources were checked; the archive date is editorial.

You find a model described as open. Can you download it? Put it in an app? Run it without sending your documents to another company? Change it? Find out what went into training it?

Those are different questions. The word “open” often gets used as though it answers all of them.

Before comparing hardware for running AI, it helps to understand what you would actually be running, what you have permission to do with it, and what you would still depend on someone else to provide.

Start with what is available

A model's weights are learned parameters. Software uses them to run a trained model without repeating its training. A weights download need not include the training data or training code. OSI separates those materials.

Here is the breakdown I would use when reading a release:

Release materials
What is available?	What I would check next
A chat website or API	Can I only use the provider's service, or can I download the model too?
Downloadable weights	Which exact release and license am I getting?
Inference code	Can I inspect and change the software that runs it?
Training code and configuration	Does this cover the actual training process, or just a sample fine-tuning script?
Training data or detailed data information	Can I examine where the data came from and how it was selected?

This is a checklist, not a ranking of model quality. A more transparent release can still be a poor fit for a particular task.

Open weights and open source

I use open weights here as a practical description of a model whose trained weights are available to download. It is not a promise that the license permits every use.

The Open Source Initiative's Open Source AI Definition 1.0 requires freedoms to use, study, modify and share, together with parameters, code and detailed training-data information. It addresses unshareable and third-party data without requiring redistribution of every raw training record. Read the definition.

If a release calls itself open source, I check its materials and terms against that named standard.

Examples from three model releases

These are specific examples, not a list of the newest or best models. The distinctions remain useful when a newer release arrives.

Qwen3-8B has an official model card identifying its license as Apache 2.0 and providing instructions for running it. That gives us something concrete to inspect. It does not establish, by itself, that every part of the training process meets the OSI AI definition. Qwen3-8B model card.

Llama 3.3 uses its own Community License. Among its conditions are attribution requirements, an incorporated acceptable-use policy, and additional commercial terms for organizations meeting a specified very-large-user-base threshold. “I can download it” is therefore an incomplete description of the permissions. Read the license for the exact version rather than assuming every release in a family has identical terms. Llama 3.3 license.

OLMo 2's 7B release shows another kind of information worth looking for. Its model card links training datasets and describes code, checkpoints, and training details, rather than presenting only a finished chat model. That makes the development process more available for study. It is not a claim here that I have independently reproduced the training or audited every released component. OLMo 2 model card.

A license also has scope. An Apache-licensed application does not automatically give its model weights, bundled datasets, or other dependencies the same license. Apache 2.0 itself has conditions, including obligations concerning license copies, notices, and modifications when redistributing covered work. “Permissive” still requires reading the terms. This is a practical explanation, not legal advice for a particular product. Apache 2.0 license.

Local and hosted processing

A downloadable model can run on your computer, on a server you manage, or through someone else's hosted service. Choosing an open-weight model does not necessarily mean choosing local processing.

To know where documents go, I need to check more than the interface. A local interface may still call a remote model, use web search, or connect to another service. Check the whole path from the document to the answer.

Ollama provides a useful concrete example: its documentation distinguishes local and cloud-hosted models, and documents a local-only mode that disables its cloud models and web search. Its privacy explanation also distinguishes local processing from its hosted service. The presence of the app on your computer is not enough to establish which route a particular request takes. Ollama FAQ.

Running locally can give you more direct control over the model version and the processing environment. You also take on the work of maintaining that environment. There is hardware, storage, power, updates, and troubleshooting to consider. A downloadable model does not come with an unlimited supply of computing resources.

Trying a model on a small task

For a first experiment, I would keep the task small and the data non-sensitive.

Pick the task before the model. Try summarizing a public document or rewriting a sample paragraph. Decide what a useful result would look like.
Choose an exact model release. Read its official model card and license. Check the runner's compatibility and memory guidance rather than assuming your computer can run every available model.
Choose where it will run. If the point is local processing, select a local model and verify that configuration. Ollama's official quickstart distinguishes local selection from cloud setup; use the current instructions rather than an old command copied from a tutorial.
Try several examples and inspect the answers. Include an example where the correct response is to acknowledge missing information. Fluency makes an answer easy to read, not necessarily reliable.
Keep a small record. Note the exact model, any quantized variant, the runner version, relevant settings, and the task. Otherwise, a later comparison may accidentally compare different setups.

For comparisons between models, I also want to check what the benchmark charts measure.

For an application, I would add a second checklist: redistribution terms, access controls, logging, update management, and how the app behaves when the model fails. A successful chat experiment is an encouraging start, not a completed production review.

Ask for the useful kind of openness

Someone studying training needs different materials from someone trying to summarize documents without sending them off a laptop. A developer distributing a product has another set of questions about permissions and dependencies.

That is why I would rather ask “what can I inspect, run, change, and ship with this release?” than stop at “is it open?”