← back to linel

Licences & attribution

What linel is built on, and what we're required or choose to credit.

Text encoder

Built with Meta Llama 3

linel's text encoder is distilled via LLM2Vec from Meta's Llama 3. Meta's licence requires this exact line to be displayed prominently on any product or service built with Llama materials or derivatives of them — it appears here and in the site footer.

Two encoder versions are currently selectable when generating:

VersionStatusNotes
v2Current, defaultDistilled from 5,000 prompts. Mean divergence 0.515 m overall / 0.646 m on the hardest split, measured against the full Llama-3 teacher on 27 held-out pairs.
v1PreviousDistilled from a smaller prompt set. Mean divergence 1.34 m against the same teacher. Kept selectable for comparison, not because it's still recommended.

Motion model

Motion model: NVIDIA Kimodo (SOMA-RP v1.1)

NVIDIA's own attribution terms apply to distributing the weights, which a hosted service like linel doesn't do — displaying this isn't strictly compelled, but it's what actually generates every clip, so we say so.

Encoder responses

Every generation carries this attribution back with it — see the attribution field in the API response, or the label shown next to a clip in the app. It is not hardcoded here or anywhere in the front end: if the model or encoder changes, what's shown follows automatically.

Questions about licensing or attribution can be sent to the service operator directly.