OpenRoots Model License
This is a plain language summary, written to be read. It is not the licence and it carries no legal force. The licence is the legal code, and where the two differ the legal code governs.
- Canonical
- openroots.org/licenses/orm/2.3
- Effective
- 2026-08-27
- Term
- Permanent
- Free below
- USD 20,000,000
You can
- Run inference commercially at any size
- Fine-tune and adapt the weights
- Distribute your fine-tunes
- Train on the model's own outputs
- Deploy in any country
You must
- State the base model and this licence on every derivative
- Carry the licence version forward to your fine-tunes
- Pay the Canopy royalty above the revenue threshold
- Publish what changed when you release an adapter
You cannot
- Sell, sublicense, redistribute, port, repackage, or host it as a competing product
- Claim the weights are open source if you altered these terms
- Impose a user-count ceiling on anyone downstream
- Exclude a region that this licence does not exclude
- Distil it into a closed model without a Compute licence
Which tier applies to you.
You do not choose a tier. It follows from your revenue and from what you are doing with the work.
Root
Freeno fee, no reporting
No Canopy royalty or report. The competing-offering, AI-training, provenance, and termination conditions still apply to every licensee.
- Keep the notice, the licence, and the version on every copy
- Do not sell, repackage, host, port, or redistribute the work itself as a competing product
- Disclose when a contribution or bug report was written with AI help
Canopy
0.5%of revenue above the threshold, capped at 250,000 USD per year
The same rights, plus a share of the revenue the work actually produced above the threshold. Only the excess is charged, so crossing the line costs almost nothing and the bill grows with the business rather than arriving whole. Self-reported quarterly in good faith. The rate cannot be negotiated by anyone, so a company adopting today gets the terms the largest adopter already has.
- Everything Root requires
- File a quarterly report and remit 0.5% of attributable revenue above the threshold
- Obtain a separate agreement before offering it as a hosted substitute
Compute
Per runpaid before the training run
Training is not covered by Root or Canopy. Human reading and ordinary search indexing are untouched. Training without this licence sits outside the licence entirely, which makes it infringement rather than a lesser breach of terms.
- Everything your base tier requires
- Pay the Compute Fee before or during the training run
- Carry provenance disclosure on outputs reproducing the work word for word
Where the uncertainty is, named by section.
Six questions counsel has not answered, published with the text and listed by section, so you can see the soft ground before you adopt anything. Most licences carry the same open questions and simply never print them.
- 01
Are model weights copyrightable at all?
Whether a trained parameter set attracts copyright protection is unsettled in most jurisdictions. If it does not, this instrument operates as a contract rather than as a licence, which changes both remedies and the position of a downstream recipient who never agreed to it.
- 02
Does Section 8.3 bind a party that never accepted the Licence?
A downstream recipient who obtains weights from an intermediary may not have accepted these terms, which limits enforceability of the prohibition against adding restrictions.
- 03
Is distillation a Derivative Work or an independent creation?
A student model trained on a teacher's outputs may not reproduce any protected expression, in which case Section 1.12 may not reach it despite Section 6 addressing it directly.
- 04
How does the European Union AI Act interact with this Licence?
Obligations attaching to general purpose AI models under that Act fall on the provider. Where a Licensee fine-tunes and distributes, whether it becomes a provider for that purpose affects duties this Licence does not address.
- 05
Can permanent source-available model terms be enforced against downstream distillation?
The current text does not rely on a conversion date. Counsel should test whether the model restrictions remain enforceable when a downstream party claims it used outputs, embeddings, adapters, or distilled weights rather than the original artifact.
- 06
What happens where training data itself was unlicensed?
This Licence governs the weights. It cannot cure a defect in the rights to the material the model was trained on, and a Licensor should not be understood to warrant otherwise.