U

UnimeType Transliteration 2.0 (v2.1) Playground

Protocol: unimetype-convert-short-v3 zh · Pinyin ja · Romaji hi · Hinglish ar · Arabizi
⌨️ Download UnimeType Keyboard App 📊 Technical Report & LM Studio Guide

1. Input & Protocol Request Builder Long-Sentence (60–240 chars)

Verified Benchmark Presets (Click to inspect)
Romanized Input (Code-switching, @handles, URLs & numbers preserved)
Operation
Rejected Candidates (for reconvert)
Compact JSON User Message (unimetype-convert-short-v3) 


        
        

          

            ⚡ Live Browser Inference (WebAssembly):
            Click to load Transliteration-2.0-0.8B (Q4_K_M · 529 MB) directly in your browser
          

          

            
          

          

            

          

        

      


      
      

        

          2. Model Output Comparison (v2.1 Long-Sentence Upgrade)
          ✓ Verified Benchmark Output
        


        

          

            UnimeType/Transliteration-2.0-4B (v2.1) · Best Accuracy (88.4% short / 93.2% long char)
            ~190 ms
          

          

        


        

          

            UnimeType/Transliteration-2.0-0.8B (v2.1) · Ultra-Light (<600 MB RAM · 72.7% short)
            ~85 ms
          

          

        


        

        

          Run Locally via LM Studio or mlx_lm (Zero System Prompt Required)
          
        

        

      

    


    
    

      

        

          Type In-Place Across Every App with UnimeType
          No Copy-Paste Needed
        

        

          
            🍎 macOS
            Native IMK + Local MLX
          
          
            🪟 Windows
            TSF Input Method
          
          
            🐧 Linux
            Desktop AT-SPI Client
          
          
            📱 iOS
            App Store Keyboard
          
          
            🤖 Android
            Google Play Keyboard
          
          
            🌐 ChromeOS
            IME Extension
          
        

      


      

        

          v2.1 Benchmark Summary (414 Short + 160 Multi-Clause Long)
          View All 6 Repos →
        

        

          
            

              Benchmark Metric
              v1.0 (4B)
              DeepSeek V4.1 Flash
              v2.1 (0.8B)
              v2.1 (4B)
            

          
          
            

              414-Case Short & Edge Strict Pass
              55.6%
              70.5%
              72.7%
              88.4%
            

            

              160-Case Multi-Clause Char Accuracy
              —
              —
              76.5%
              93.2%
            

            

              Chinese Pinyin (zh · 207 cases)
              72.9%
              82.6%
              81.2%
              94.2%
            

            

              Japanese Romaji (ja · 52 cases)
              28.8%
              71.2%
              42.3%
              76.9%
            

            

              Hindi Hinglish (hi · 71 cases)
              57.7%
              69.0%
              71.8%
              85.9%
            

            

              Arabic Arabizi (ar · 84 cases)
              27.4%
              41.7%
              71.4%
              83.3%