Phi-3-mini on raw WebGPU
Runs in your browser on hand-written WGSL kernels. Weights download once and cache locally; prompts and generation stay on this machine.
~2 GB
weights
~70 t/s
M2 Max
—
context
Pick a different model →
Not now
Preparing Zero-TVM
Starting…
0%
Details 

      

      

    

  


  
  

    

      
        
      
      

        
        
hand-written WGSL · paged KV · WebGPU

        
      

      

      
      

        
        
      

      
      Waiting
    


    

      

        

          
          
Chat with Phi-3-mini

          

            q4f16_1 · 3.8 B
            4 K context
            Runs on your GPU · WebGPU
            On-device only
          

          
Running on your own GPU through hand-written WGSL kernels. Prompts, tokens and the KV cache stay in this tab.

          

            
            
            
            
          

        

        

      

      
    


    

      

        
        
        
      

      

        Enter to send · Shift+Enter for new line ·
        Zero TVM · 10 WGSL kernel roles