SigLIP 2 Base

SigLIP 2 Base is a multilingual vision-language encoder for zero-shot image classification, image-text retrieval, and dense prediction tasks, hosted as a managed API on gigarouter.

Architecture and Capabilities

The model (ViT-B, 86M parameters) extends the SigLIP pretraining objective with captioning-based pretraining, self-supervised losses (self-distillation, masked prediction), and online data curation. It uses the Gemma tokenizer with a 256k vocabulary. The NaFlex variant preserves the input's native aspect ratio and supports multiple resolutions, requiring adapted preprocessing not compatible with standard SigLIP inference code.

Training Details

Pre-trained on the WebLI dataset (Chen et al., 2023) with a diverse data mixture incorporating de-biasing techniques for improved multilingual understanding and fairness. Training used up to 2048 TPU-v5e chips.

Benchmark Results

At 224px resolution, SigLIP 2 Base achieves:

• 78.2% ImageNet zero-shot accuracy
• 52.1 COCO text-to-image recall
• 68.9 COCO image-to-text recall

The model outperforms its SigLIP counterpart at the same scale across zero-shot classification, retrieval, and transfer performance for Vision-Language Models (VLMs), with significant gains on localization and dense prediction tasks.

Model Variants

SigLIP 2 is released in four sizes: ViT-B (86M), L (303M), So400m (400M), and g (1B). This base model is the smallest, optimized for efficient inference.