SigLIP 2

SigLIP 2 is a family of vision-language encoders published by Google DeepMind on 20 February 2025. It does not answer questions and does not generate text. It maps an image and a piece of text into the same vector space, which is enough for two jobs that turn up everywhere: zero-shot image classification, where the candidate labels are simply written out as text at inference time, and image-text retrieval. Its third and least visible job matters most — it is the standard vision tower bolted onto multimodal language models, Google's own MedGemma among them. The second generation keeps the sigmoid loss that gave the line its name and adds three training ideas on top: a decoder loss, a global-local and masked prediction loss, and adaptability to aspect ratio and resolution. The point of the additions is exactly what the paper's title promises — better semantic understanding, better localisation and better dense features, meaning the encoder is more useful when something has to be found in a picture rather than merely recognised. Pre-training runs on the WebLI multilingual image-text dataset, on up to 2,048 TPU v5e chips. The release ships as a grid rather than a single model: four sizes — base at 375 million parameters, large at 882 million, so400m at 1.14 billion and giant-opt at 1.87 billion — each at resolutions from 224 to 512 pixels, plus a separate NaFlex line that keeps the native aspect ratio of the image instead of squashing it into a square. There are JAX checkpoints alongside the PyTorch ones. Traffic concentrates at the bottom of the range: the 256-pixel base model alone is downloaded 3.89 million times a month, more than any other Google model in this catalogue, because it is the cheapest encoder that is good enough to build on. Everything is Apache 2.0, with no gate and no terms to accept, which is the practical reason why so many third-party models are built on it. Google publishes the benchmark comparison against SigLIP 1 only inside the paper and as an image on the model card, so this catalogue does not quote those figures.

▸Videos

No videos yet.

▸Reviews