Updated · Sources & corrections
In this guide
A local model can be useful without being a fully open-source system. To evaluate it, keep three questions separate: can you obtain the files, what may you do with them, and what information lets you inspect or modify the system? This guide provides a practical release review, with two licence examples rather than a blanket verdict on entire model families.
Open weights and open source answer different questions
The Open Source Initiative’s Open Source AI Definition 1.0 describes freedoms to use, study, modify and share, supported by data information, code and parameters under the specified open terms. Its data requirement includes provenance, selection and processing details and listings of obtainable datasets. It does not simply require publishing every raw training record.
Downloadable weights alone do not establish that a release meets those requirements. Conversely, a permissive licence on weights does not prove that training code and data information were released. Assess the actual components instead of using the publisher’s headline as a checklist.
| Question | Evidence to inspect | What it does not establish |
|---|---|---|
| Can I run this checkpoint? | Weight files, architecture, tokenizer and runtime support | Permission for every application or distribution method |
| What reuse is permitted? | Exact licence text, notices and incorporated policies | Availability of the training recipe |
| Can I study and modify the system? | Data information, training/inference code and parameters against the definition being used | A guarantee that reproduction will match every result |
Start with the exact artifact
Record the publisher, repository, checkpoint revision and downloaded file. A family name is insufficient: a base checkpoint, an instruction-tuned release and a third-party quantization may carry different accompanying files. Trace a conversion back to its base model and retain both sets of relevant notices.
Hugging Face supports licence identifiers in repository-card metadata. Treat that badge as a route to the terms, then read the actual file. If the badge and file disagree, record the discrepancy and seek clarification before relying on a permission.
Two examples: Apache 2.0 and Llama 3.1
Apache 2.0 is permissive, with redistribution duties
Apache License 2.0 grants broad reuse rights subject to its conditions. Section 4 requires a licence copy for recipients, notices on modified files, preservation of relevant notices and handling of an included NOTICE file. Section 6 limits trademark rights. A weight release under Apache 2.0 therefore still needs an attribution and distribution review; a permissive label does not mean there are no obligations.
Llama 3.1 has a custom licence
The Llama 3.1 licence supplied with Meta’s 8B Instruct checkpoint includes conditions on redistribution, attribution and certain model naming, and incorporates an acceptable-use policy. Section 2 refers to more than 700 million monthly active users, measured for the relevant products/services and affiliates in the preceding calendar month on the version release date. This is a user threshold, not a revenue threshold. Read the full clause for applicability; do not transfer this summary to a different Llama version.
Neither example is an approval of a particular deployment. They show why a checkpoint-specific record is more useful than assuming that all downloadable models have identical terms.
A checkpoint review you can reuse
Before reviewing, identify your planned activity: internal inference, a hosted service, redistribution of weights, fine-tuning, or training with outputs. These activities need separate answers. Use the following record as a worksheet, not as evidence that a review has already passed.
Checkpoint review — illustrative blank record
Publisher / repository:
Revision / file / checksum:
Base checkpoint and conversion source:
Licence URL and saved copy date:
Linked policies and version/date:
Planned activity and distribution method:
Relevant permission clauses:
Restrictions / thresholds to investigate:
Licence, attribution and NOTICE delivery plan:
Modified-file or derivative naming requirements:
Training code and data information available:
Unresolved questions and reviewer:
Decision / evidence / next review trigger:Store the record alongside your deployment manifest so a model upgrade triggers a fresh check. Keep software runtime dependencies in a separate component list: their licence does not replace the weights’ terms. For the runtime side, see our Ollama setup and configuration guide and model storage guide.
When the evidence is incomplete
| Finding | Next step |
|---|---|
| No licence file or unclear permission | Ask the publisher to clarify; do not infer permission from download availability. |
| Custom or research-only terms | Map the intended activity to the actual text and seek qualified review where needed. |
| Third-party converted checkpoint | Trace the base release and compare its terms and notices with the conversion repository. |
| Training information missing | Describe that specific transparency gap; do not conclude that the model cannot run or that every use is forbidden. |
| Model, revision or deployment changes | Reopen the record instead of carrying forward an old approval automatically. |
Your next step is to complete the record for one candidate checkpoint before investing in integration. Then evaluate runtime behaviour on your own tasks; licence review and technical evaluation resolve different uncertainties.
Sources and verification scope
AI Models News reviewed the linked primary documentation on 24 October 2026. This is a documentation-based explanation, not legal advice or a certification of any model’s compliance. No model execution, benchmark or legal approval is claimed. The worksheet is illustrative. This update corrects the earlier user/revenue threshold wording, data-disclosure simplification and redistribution summary. Read our editorial policy.
When adapting a licensed checkpoint, separate training configuration from reuse permission. Our LoRA versus QLoRA guide explains frozen weights, quantization and how to compare training memory without treating weight-size arithmetic as a GPU benchmark.
For a concrete way to record the file version you reviewed, follow our pinned Hugging Face download and checksum example. It checks one public configuration file and a changed local copy, while keeping file identity separate from licence permission and publisher authenticity.