← Back to Hub

Error loading visualization








  

    
VLM Parameters

    
ModelPaliGemma-3B

    
LLM BackboneGemma-2B

    
Vision EncoderSigLIP-So400M

    
Embed Dim2048

    
Decoder Layers18

    
Attention Heads16

    
WeightsFrozen

  

  

    
Processing Progress

    

  

  

    
Token Sequence (23 shown)

    

  





  

    
Legend

    
Text Prompt / Language

    
PaliGemma Decoder

    
Visual Tokens (SigLIP)

    
Cross-Attention Beams

    
Data Flow

    
Fused Embeddings

  







  

    
Ready

    
Press Play to see how PaliGemma combines visual tokens from SigLIP with language tokens to create fused vision-language embeddings — Step 2 of the pi0 architecture.

  

  

    
    
    
    
Speed1x