SigLIP 2 Base (patch 16, 224 px), 16-bit copy

A 16-bit (float16) copy of Google's SigLIP 2 Base, an image and text encoder that puts pictures and words in one 768-wide space. At half the size of the original 32-bit file, it is the reference embedder for UI guides in Local Model Access (LMA) - created by artist Josh Hughes.

A site that ships a UI guide embeds pictures of its interface with this file. A visitor's device embeds their screenshot or question with the same file, on the device, and finds the guide entry that answers "how do I use this?". It is the example LMA's tools and tests use, not a requirement: each site chooses its own embedder and names it by identity in its guide.

Files

File What it is
model.safetensors The weights in float16, with tokenizer.model inside (a file:tokenizer.model tensor): 754,665,695 bytes, sha256 0588cc7dfd605c6c50a6a08355581cac8e2b02eff3697e17373b67059f58ba9a. This one file is all a visitor downloads
config.json, preprocessor_config.json Unchanged from the source
tokenizer.model, tokenizer.json, tokenizer_config.json, special_tokens_map.json Unchanged from the source (Gemma's SentencePiece vocabulary)
LICENSE, NOTICE Apache-2.0, and what was changed

Source

  • Repository: google/siglip2-base-patch16-224
  • Revision: 75de2d55ec2d0b4efc50b3e9ad70dba96a7b2fa2
  • File: model.safetensors, 1,500,800,904 bytes, sha256 612923381c76ec5a9bed335d1c48827e3f2e506ac31b044b63b2031fadee6a0b (float32)

How it was converted

  • Weights: every float32 tensor was rounded to float16 (round to nearest, ties to even) with LMA's tools/f16_copy.py, one tensor at a time. The 408 tensor names, shapes and order are unchanged, and nothing else was changed.
  • Metadata: the file's __metadata__ names the source and the conversion.
  • The tokenizer inside: tokenizer.model is carried in the file as a 1-D U8 tensor named file:tokenizer.model (LMA's SPEC.md §2.1), so the model is one file to download and pick. transformers loads the file as before and lists that tensor as unused.
  • Accuracy: the largest relative rounding error, over values above 1e-4, is 4.9e-4. Image and text embeddings from this file match the 32-bit original at a cosine similarity of 0.99998 or better on the pictures and questions tested.
  • Companion files: copied as they are.

Use

With transformers (computed in float32 on load):

import torch
from transformers import AutoModel, AutoProcessor
model = AutoModel.from_pretrained("JoshHughes/siglip2-base-patch16-224-f16", dtype=torch.float32).eval()
processor = AutoProcessor.from_pretrained("JoshHughes/siglip2-base-patch16-224-f16")

The model was trained on lowercased text (do_lower_case in tokenizer_config.json), so lowercase questions before tokenising, and pad to 64 tokens (padding="max_length", max_length=64), as SigLIP models expect.

In a browser, LMA's embedder (lma/embed/) runs this file on WebGPU or the CPU. It reads the file a slice at a time from the visitor's own copy, and reads only the token rows a question uses.

Licence and credit

SigLIP 2 is by Google: Tschannen et al., SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features, 2025. The original weights are released under the Apache License 2.0, and so is this modified copy (LICENSE). NOTICE says what changed.

@misc{tschannen2025siglip2multilingualvisionlanguage,
      title={SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features},
      author={Michael Tschannen and Alexey Gritsenko and Xiao Wang and Muhammad Ferjad Naeem and Ibrahim Alabdulmohsin and Nikhil Parthasarathy and Talfan Evans and Lucas Beyer and Ye Xia and Basil Mustafa and Olivier Hénaff and Jeremiah Harmsen and Andreas Steiner and Xiaohua Zhai},
      year={2025},
      eprint={2502.14786},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2502.14786},
}
Downloads last month
28
Safetensors
Model size
0.4B params
Tensor type
F16
·
U8
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for JoshHughes/siglip2-base-patch16-224-f16

Finetuned
(155)
this model

Paper for JoshHughes/siglip2-base-patch16-224-f16