Open Weight vs Open Source AI: What Changes

Open Weight Is Not the Same as Open Source

An open-weight model makes learned parameters available under stated terms. Open source AI is a broader claim about freedoms, components, and preferred forms for modification. Neither label should replace a review of the exact artifacts and license.

The safest public language describes each access dimension separately.

flowchart TD
    A["AI release"] --> B["Weights"]
    A --> C["Architecture and inference code"]
    A --> D["Training and data information"]
    A --> E["Documentation and evaluations"]
    A --> F["License rights and restrictions"]
    B --> G["Practical control"]
    C --> G
    D --> G
    E --> G
    F --> G

What open weight means

Weights are learned parameters used by a model architecture to produce outputs. When a publisher provides them, an adopter may be able to download and run the model in a compatible environment.

What the adopter may legally do depends on the governing terms.

Weight access does not automatically provide the training dataset, data-processing code, training code, optimizer state, evaluation data, full reproducibility, unrestricted redistribution, or permission for every use.

It also does not guarantee that a particular conversion, quantization, or mirror came from the publisher.

What open source AI means

The Open Source AI Definition 1.0 centers freedoms to use, study, modify, and share. It also describes access to preferred forms for modification, including data information, code, and parameters.

That definition provides a concrete standard. Other speakers may use "open source" differently, which is why a public claim should name the definition being applied.

The point is not to police casual conversation. The point is to prevent a procurement, product, or compliance decision from resting on an undefined marketing term.

A useful access record

For each model, record the publisher, family, exact version, base or instruction-tuned form, weight source, file names, hashes, architecture, tokenizer, prompt format, inference code, training code, data information, model card, evaluations, safeguards, license, acceptable-use policy, and update history.

Then record the allowed and restricted actions for the intended use. Include internal operation, commercial use, modification, fine-tuning, distillation, output use, redistribution, hosting, attribution, naming, scale, territory, and termination where applicable.

Unknown does not mean allowed. It means unresolved.

Llama 3.1 as an example

Meta's Llama 3.1 model card documents the official family, architecture, intended use, supported languages, evaluations, and safety information.

The Llama 3.1 Community License defines the governing grant and conditions. It includes attribution, redistribution, naming, acceptable-use, and additional commercial boundaries.

That combination supports describing Llama 3.1 as an open-weight release under a custom community license. It should not be described as public domain or unrestricted free use.

Qualified legal review is required when the terms affect a real decision.

Access does not equal reproducibility

A team may be able to run a released checkpoint without being able to reconstruct how it was produced.

Reproducibility can depend on the source data, processing and filtering logic, training code, hyperparameters, distributed system, random seeds, post-training data, evaluation data, and exact dependencies.

The official Llama 3 paper provides substantial technical evidence. It does not make every training object or operational detail available.

That is not an accusation. It is a boundary in what the public evidence supports.

Practical control also varies

An adopter may have weights but still rely on a third-party runtime, cloud service, conversion, tokenizer, safeguard, or model host.

Self-hosting can increase control over infrastructure and data flow. It also transfers patching, capacity, monitoring, access, incidents, deletion, and continuity to the adopter.

A hosted endpoint for an open-weight model may offer less local control than a self-hosted deployment. A closed model service may offer stronger managed controls for a particular team.

The label alone does not decide the operating model.

Use precise public language

Prefer sentences such as:

"The publisher provides model weights under a custom license."

"The repository includes inference code and a model card."

"Training-data information is described at a high level, but the document-level corpus is not published."

"This deployment uses a third-party quantized conversion whose provenance requires separate verification."

Those sentences tell a reader what is actually available.

Episode 27 used open-source language broadly because the release felt like a major expansion of access. The corrected lesson is stronger: openness is multidimensional, and the model decision should show its work.

Use [[How to Read and Operationalize an AI Model License]] for the terms-to-controls method. Use [[Llama 3.1 Technical Release Record]] for the versioned Llama example.

This explainer was developed with AI assistance from E027, the Open Source Initiative, official Llama sources, and the linked taxonomy. Dalton Anderson remains the author. Editorial, legal, license, technical, source, and founder review are mandatory before publication or use in a decision. Publication is not authorized.