siglip2-so400m-patch16-naflex
by google
SigLIP 2 at native aspect ratio and variable resolution
google/siglip2-so400m-patch16-naflexOverview
SigLIP 2 replaces CLIP's softmax contrastive loss with a pairwise sigmoid loss, which removes the need for a global normalization across the batch and trains stably at large scale. The NaFlex variant adds the property that matters most for real documents and photographs: it handles native aspect ratios and variable sequence lengths instead of forcing every input into a fixed square.
Squashing a 16:9 video frame or a portrait page into a square distorts the spatial relationships a retrieval model is trying to encode. NaFlex avoids that, which is why it tends to do better on documents, screenshots and wide video frames than fixed-resolution encoders.
At 1.14B parameters this is a large encoder, so it usually belongs in a reranking or high-precision stage rather than running over every frame.
Architecture
SigLIP 2 so400m (shape-optimized 400M-class vision tower) with 16x16 patches and NaFlex support for native aspect ratio and variable sequence length. Trained with a pairwise sigmoid contrastive loss. 1,135,670,962 parameters total per the model card.
Key Capabilities
- •Native aspect ratio input, with no forced square resize
- •Variable sequence length, so resolution can scale with content
- •Sigmoid contrastive loss, stable at large batch scale
- •Zero-shot classification and cross-modal retrieval
Use Cases on Mixpeek
- •Document and screenshot retrieval where aspect ratio carries layout meaning
- •Wide video frame embedding without distortion from square cropping
- •High-precision reranking over candidates recalled by a smaller encoder
- •Cross-modal search where fine visual detail decides relevance
Tags
Use siglip2-so400m-patch16-naflex on Mixpeek
Build multimodal processing pipelines with this model and others. Extract features, run inference, and set up retrieval in Mixpeek Studio.
Open StudioHow It Runs on Mixpeek
On Mixpeek, siglip2-so400m-patch16-naflex runs as a managed extractor inside a processing pipeline. Point a bucket of zero shot image classification data at it, and Mixpeek handles GPU provisioning, batching, retries, and writing the outputs into a vector store you can query.
Extractor outputs land in the Mixpeek Vector Store (MVS), where you can combine them with retrieval, reranking, and filter stages to build end-to-end search and agent-perception pipelines, no model-serving infrastructure to maintain.
Explore on Mixpeek
Compare alternatives in this category
Hand-picked tools & platforms compared
Deep-dive technical guide
See how Mixpeek runs models as extractors
Store & search embeddings at scale
Usage-based pricing for pipelines
Compare models, APIs & infrastructure
Specification
View on HuggingFace
See model card, files, and community discussion