Licensing and origin
Published by Vox (Tokyo, Japan) under the v13s namespace. Tancho Generator is a derivative of NVIDIA Cosmos3-Edge, pinned revision a9d944e2c6a1bf9f48b92ad16348e70c5f1836ba. Copyright and origin notices in the original retained artifacts are preserved. The pinned upstream model card and available license/notice files are copied under licenses/upstream/ for attribution; its full-model capabilities are not claims about this Tancho release. Model weights and upstream model artifacts are distributed under the accompanying licenses/LICENSE.OpenMDW-1.1. Original Vox code, documentation, and synthetic evaluation records use LICENSE-CODE (MIT); that license does not replace third-party terms.
The pinned NVIDIA Cosmos Framework is a separate dependency, not bundled source software. Install it under its own terms. The exported processor/tokenizer/configuration assets retain their original bytes. The checkpoint provenance files refer to the pinned upstream base snapshot; trained-export file identity is established by release-manifest.json and the checkpoint SHA-256 records.
No GSI footage, extracted frames, client materials, credentials, or complete internal run archives are distributed. Public GSI source imagery informed training; see SOURCES.md. The packaged evidence contains synthetic numerical geometry, image hashes and saved generated text, with no images. These are developer-created records, not official GSI assessments. No NVIDIA, Nebius, or GSI endorsement is asserted.
The dependency file deploy/nebius/edge-parent40-expected-weights.json supplies a historical fingerprint for checking the frozen vision encoder in the audited overlay helper. It is not a parent checkpoint and is not loaded as trained weights. The two released models were independently fine-tuned from the pretrained base.