SigLIP 2 -- so400m16-256
sigmoid-loss image-text pretraining, two towers in one shared space

Copyright 2025 The Google Research Authors.
Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem,
Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer,
Ye Xia, Basil Mustafa, Olivier Henaff, Jeremiah Harmsen, Andreas Steiner,
Xiaohua Zhai.

Source:  https://huggingface.co/google/siglip2-so400m-patch16-256
Project: https://github.com/google-research/big_vision
Paper:   SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic
         Understanding, Localization, and Dense Features (2025), arXiv:2502.14786
Licence: Apache-2.0 (full text in the LICENSE file beside this one)

The authors state of their SigLIP releases: these models are not official Google
products and were trained and released for research purposes.

Google's own release is JAX/Flax. This file derives from the PyTorch conversion Google
publishes beside it on Hugging Face, under the same terms. It is not the byte stream
Google serves: tools/fetch/siglip2.py verifies the safetensors against the sha256 the
Hub records, then writes the same tensors back out as an ordinary checkpoint. No tensor
is altered, renamed, cast or dropped. The source files are:
           model.safetensors  sha256 810ddc85ac019e6b2738b9130ae3602e41eb093a848a297a0ad296ca1f39dc67

The tokenizer vocabulary shipped inside mozo derives from tokenizer.model in this same
repository. It is Gemma's vocabulary, and Gemma's own weights are published under the
Gemma Terms of Use behind a gate -- but Google publishes this copy here, ungated, in a
repository licensed Apache-2.0. That is an affirmative grant by the rights holder on the
artifact actually taken, not an inference from silence.