# siglip2-base-patch16-naflex by Google: Open-Weight Model Source: https://savrn.com/models/siglip2-base-patch16-naflex Markdown alternate of the page above; the site index is https://savrn.com/llms.txt --- ## Runs On What it takes to serve siglip2-base-patch16-naflex (375M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them. | Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits | | --- | --- | --- | --- | --- | --- | | 16-bit | 0.8 GB | 0.9 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | | 8-bit | 0.4 GB | 0.5 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | | 4-bit | 0.2 GB | 0.2 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the [SAVRN Index](https://savrn.com/ai-index/pricing/gpus), read Oct 23, 2026. [siglip2-base-patch16-naflex on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/siglip2-base-patch16-naflex/gpus) ## SAVRN's Notes on siglip2-base-patch16-naflex Google built this encoder for three jobs: sorting images against text labels with no task-specific training, matching images to captions for retrieval, and serving as the vision tower inside a larger vision-language model. At 375 million parameters it needs 0.9 GB of memory at 16-bit precision, so hardware choice is about what else is on the card. The cheapest card in our table, one MI300X with 192 GB at $1.85 an hour on demand, carries it alongside the language model it feeds. Apache 2.0 clears commercial use, modification and redistribution with the notices kept. Three papers describe the method, arXiv 2502.14786, 2303.15343 and 2209.06794, and the first titles it a multilingual vision-language encoder, which matters if your labels are not in English. The release is February 2025, so confirm your inference stack supports the siglip2 model type before ordering hardware. SAVRN Research, 2026-09-18 ## Model Card By Google, published under apache-2.0, revision b53b807d3a2d. ### SigLIP 2 Base [SigLIP 2](https://savrn.com/papers/siglip-2-multilingual-vision-language-encoders-with-improved-semantic-understanding-locali) extends the pretraining objective of [SigLIP](https://savrn.com/papers/sigmoid-loss-for-language-image-pre-training) with prior, independently developed techniques into a unified recipe, for improved semantic understanding, localization, and dense features. ### Intended uses You can use the raw model for tasks like zero-shot image classification and image-text retrieval, or as a vision encoder for VLMs (and other vision tasks). Here is how to use this model to perform zero-shot image classification: ``` from transformers import pipeline # load pipeline ckpt = "google/siglip2-base-patch16-naflex" image_classifier = pipeline(model=ckpt, task="zero-shot-image-classification") # load image and candidate labels url = "http://images.cocodataset.org/val2017/000000039769.jpg" candidate_labels = ["2 cats", "a plane", "a remote"] # run inference outputs = image_classifier(image, candidate_labels) print(outputs) ``` You can encode an image using the Vision Tower like so: [Read the full model card (307 words)](https://savrn.com/models/siglip2-base-patch16-naflex/card) ## Configuration Architecture Siglip2Model Vocabulary size 256,000 Stored precision float32 Model type siglip2 ## Identity and Version Repository google/siglip2-base-patch16-naflex Publisher Google Task Zero shot image classification Modality Other Library transformers Parameters 375M parameters Languages Not stated by the source Revision b53b807d3a2d5e2b3911292f2d69e5341cdc064c First published 2025-02-18 Last updated 2025-02-21 ## Files and Weights 9 files, 1.5 GB in total. The weights are 1 file totalling 1.5 GB in safetensors. Weights1 file · 1.5 GB Configuration3 files · 1.4 KB Tokenizer3 files · 38.6 MB Documentation1 file · 3.4 KB Repository1 file · 1.6 KB Every file | File | Type | Size | SHA-256 | | --- | --- | --- | --- | | model.safetensors | Weights | 1.5 GB | ac5f28bbdf92 | | config.json | Configuration | 329 B | — | | preprocessor_config.json | Configuration | 393 B | — | | special_tokens_map.json | Configuration | 636 B | — | | README.md | Documentation | 3.4 KB | — | | .gitattributes | Repository | 1.6 KB | — | | tokenizer.json | Tokenizer | 34.4 MB | 58a1696e79c9 | | tokenizer.model | Tokenizer | 4.2 MB | 61a7b147390c | | tokenizer_config.json | Tokenizer | 40.2 KB | — | ## License and Download License apache-2.0 Access Open weights, no gate Download size 1.5 GB [Download from Google](https://huggingface.co/google/siglip2-base-patch16-naflex) Released by Google through its official repository on Hugging Face. [Read the license](https://www.apache.org/licenses/LICENSE-2.0). ## Built From - Described by [arXiv:2209.06794](https://savrn.com/papers/pali-a-jointly-scaled-multilingual-language-image-model) - Described by [arXiv:2303.15343](https://savrn.com/papers/sigmoid-loss-for-language-image-pre-training) - Described by [arXiv:2502.14786](https://savrn.com/papers/siglip-2-multilingual-vision-language-encoders-with-improved-semantic-understanding-locali) ## Memory Requirements | Precision | Weights in memory | | --- | --- | | As published | 1.5 GB | | 16-bit | 0.8 GB | | 8-bit | 0.4 GB | | 4-bit | 0.2 GB | Weights only, from the published parameter count; the key-value cache and runtime add to this. ## Compare siglip2-base-patch16-naflex - [CLIP-ViT-L-14-laion2B-s32B-b82K vs siglip2-base-patch16-naflex](https://savrn.com/models/compare/clip-vit-l-14-laion2b-s32b-b82k-vs-siglip2-base-patch16-naflex) - [clip-vit-large-patch14 vs siglip2-base-patch16-naflex](https://savrn.com/models/compare/clip-vit-large-patch14-vs-siglip2-base-patch16-naflex) - [siglip2-base-patch16-256 vs siglip2-base-patch16-naflex](https://savrn.com/models/compare/siglip2-base-patch16-256-vs-siglip2-base-patch16-naflex) ## Questions About siglip2-base-patch16-naflex ### How much GPU memory does siglip2-base-patch16-naflex need? About 0.9 GB at 16-bit and 0.2 GB at 4-bit: the weights (375M parameters) plus a working margin. A long context needs more. ### What is the cheapest GPU to run siglip2-base-patch16-naflex on? At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists. ### Can I use siglip2-base-patch16-naflex commercially? Yes. siglip2-base-patch16-naflex is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors. ## Similar Models Model · Zero shot image classification ### [siglip2-base-patch16-256](https://savrn.com/models/siglip2-base-patch16-256) [Google](https://savrn.com/model-publishers/google) SigLIP 2 extends the pretraining objective of SigLIP with prior, independently developed techniques into a unified recipe, for improved semantic understanding, localization, and dense features. You can use the raw model for tasks like zero-shot image classification and image-text retrieval, or as a vision encoder for VLMs (and other vision tasks). Here is how to use this model to perform zero-shot image classification: You can encode an image using the Vision Tower like so: For more code examples, we refer to the siglip documentation. SigLIP 2 adds some clever training objectives on top of SigLIP: SigLIP 2 is pre-trained on the WebLI dataset (Chen et al., 2023). The model was trained on up… Open weights apache-2.0 375M parameters transformers [View model](https://savrn.com/models/siglip2-base-patch16-256) Model · Zero shot image classification ### [siglip2-base-patch16-224](https://savrn.com/models/siglip2-base-patch16-224) [Google](https://savrn.com/model-publishers/google) SigLIP 2 extends the pretraining objective of SigLIP with prior, independently developed techniques into a unified recipe, for improved semantic understanding, localization, and dense features. You can use the raw model for tasks like zero-shot image classification and image-text retrieval, or as a vision encoder for VLMs (and other vision tasks). Here is how to use this model to perform zero-shot image classification: You can encode an image using the Vision Tower like so: For more code examples, we refer to the siglip documentation. SigLIP 2 adds some clever training objectives on top of SigLIP: SigLIP 2 is pre-trained on the WebLI dataset (Chen et al., 2023). The model was trained on up… Open weights apache-2.0 375M parameters transformers [View model](https://savrn.com/models/siglip2-base-patch16-224) Model · Zero shot image classification ### [clip-vit-large-patch14](https://savrn.com/models/clip-vit-large-patch14) [OpenAI](https://savrn.com/model-publishers/openai) Disclaimer: The model card is taken and modified from the official CLIP repository, it can be found here. The CLIP model was developed by researchers at OpenAI to learn about what contributes to robustness in computer vision tasks. The model was also developed to test the ability of models to generalize to arbitrary image classification tasks in a zero-shot manner. It was not developed for general model deployment - to deploy models like CLIP, researchers will first need to carefully study their capabilities in relation to the specific context they’re being deployed within. January 2021 The base model uses a ViT-L/14 Transformer architecture as an image encoder and uses a masked… Open weights 428M parameters 77 tokens transformers [View model](https://savrn.com/models/clip-vit-large-patch14) Model · Zero shot image classification ### [CLIP-ViT-L-14-laion2B-s32B-b82K](https://savrn.com/models/clip-vit-l-14-laion2b-s32b-b82k) [LAION eV](https://savrn.com/model-publishers/laion) A CLIP ViT L/14 model trained with the LAION-2B English subset of LAION-5B (https://laion.ai/blog/laion-5b/) using OpenCLIP (https://github.com/mlfoundations/openclip). Model training ('babysitting') done by Ross Wightman on the JUWELS Booster supercomputer. See acknowledgements below. As per the original OpenAI CLIP model card, this model is intended as a research output for research communities. We hope that this model will enable researchers to better understand and explore zero-shot, arbitrary image classification. We also hope it can be used for interdisciplinary studies of the potential impact of such model. The OpenAI CLIP paper includes a discussion of potential downstream impacts… Open weights mit 428M parameters 77 tokens open_clip [View model](https://savrn.com/models/clip-vit-l-14-laion2b-s32b-b82k) Model · Zero shot image classification ### [siglip-base-patch16-224](https://savrn.com/models/siglip-base-patch16-224) [Google](https://savrn.com/model-publishers/google) SigLIP model pre-trained on WebLi at resolution 224x224. It was introduced in the paper Sigmoid Loss for Language Image Pre-Training by Zhai et al. and first released in this repository. Disclaimer: The team releasing SigLIP did not write a model card for this model so this model card has been written by the Hugging Face team. SigLIP is CLIP, a multimodal model, with a better loss function. The sigmoid loss operates solely on image-text pairs and does not require a global view of the pairwise similarities for normalization. This allows further scaling up the batch size, while also performing better at smaller batch sizes. A TLDR of SigLIP by one of the authors can be found here. You can use… Open weights apache-2.0 203M parameters transformers [View model](https://savrn.com/models/siglip-base-patch16-224) Model · Zero shot image classification ### [CLIP-ViT-B-32-laion2B-s34B-b79K](https://savrn.com/models/clip-vit-b-32-laion2b-s34b-b79k) [LAION eV](https://savrn.com/model-publishers/laion) A CLIP ViT-B/32 model trained with the LAION-2B English subset of LAION-5B (https://laion.ai/blog/laion-5b/) using OpenCLIP (https://github.com/mlfoundations/openclip). Model training done by Romain Beaumont on the stability.ai cluster. As per the original OpenAI CLIP model card, this model is intended as a research output for research communities. We hope that this model will enable researchers to better understand and explore zero-shot, arbitrary image classification. We also hope it can be used for interdisciplinary studies of the potential impact of such model. The OpenAI CLIP paper includes a discussion of potential downstream impacts to provide an example for this sort of analysis.… Open weights mit 151M parameters 77 tokens open_clip [View model](https://savrn.com/models/clip-vit-b-32-laion2b-s34b-b79k) ## Google [All models and datasets](https://savrn.com/model-publishers/google) ## Versions - [b53b807d3a2d](https://savrn.com/models/siglip2-base-patch16-naflex/versions/b53b807d3a2d) · current 2026-10-23 ## Explore More - [All zero shot image classification models](https://savrn.com/models/tasks/zero-shot-image-classification) - [All models under apache-2.0](https://savrn.com/models/licenses/apache-2-0) - [Model comparisons](https://savrn.com/models/comparisons) - [The model directory](https://savrn.com/models) - [Open model prices by host](https://savrn.com/ai-index/pricing/open-models) ## Source - Repository metadata, read 2026-10-19. - [Google website](https://ai.google.dev/) - [Hugging Face record](https://huggingface.co/google/siglip2-base-patch16-naflex) - [How the hub is built](https://savrn.com/model-hub/methodology)