sam3_deploy
Copyright 2025 The Meta AI Authors and The HuggingFace Team.

This product includes software derived from the SAM 3 implementation in HuggingFace
Transformers (https://github.com/huggingface/transformers, src/transformers/models/sam3),
which is licensed under the Apache License, Version 2.0; see the LICENSE file in this
directory. That implementation is co-authored by Meta AI and HuggingFace and carries an
Apache-2.0 header, which is why it -- and not facebookresearch/sam3 -- is the source this
package derives from.

The byte-pair tokenizer and its vocabulary come from OpenAI's CLIP
(https://github.com/openai/CLIP) and are MIT licensed. `assets/bpe_simple_vocab_16e6.txt.gz`
is byte-identical to the file published there:

  sha256  924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a

WEIGHTS ARE NOT APACHE-2.0
--------------------------
Unlike this code, SAM 3's published checkpoints are covered by Meta's SAM License, a
bespoke agreement that is not an OSI-approved open-source licence. It carries field-of-use
restrictions -- no military, nuclear, weapons or espionage use -- that flow through to
anyone you serve predictions to, and it binds recipients on use rather than on assent. Any
distribution must ship a copy of that agreement alongside the weights. mozo does not
redistribute them from this package.

NOT INCLUDED
------------
The video path is not part of this package: memory attention, the memory encoder, temporal
disambiguation and the tracklet bookkeeping that drives them are all left behind, along
with the training code and the evaluation harness. See PROVENANCE.md.
