upd

Author: gyt1145028706

.gitignore

@@ -46,4 +46,5 @@ dev/*
 
 .codex
 .claude/
-.github/
+.github/
+dev/*

assets/moss-tts-nano.png

@@ -91,7 +91,7 @@ <h1>MOSS-TTS-Nano</h1>
         <div class="meta-grid">
           <div class="meta-item">
             <p class="meta-label">Parameters</p>
-            <p class="meta-value">100 M</p>
+            <p class="meta-value">~100 M</p>
           </div>
           <div class="meta-item">
             <p class="meta-label">Audio Supported</p>
@@ -150,7 +150,7 @@ <h2>Key Features</h2>
           <div class="feature-item">
             <p class="feature-label">Audio Quality</p>
             <p class="feature-value">48 kHz Stereo</p>
-            <p class="feature-desc">Native 2-channel output at full 48 kHz sample rate.</p>
+            <p class="feature-desc">Native 2-channel input and output at full 48 kHz sample rate.</p>
           </div>
           <div class="feature-item">
             <p class="feature-label">Languages</p>
@@ -202,11 +202,11 @@ <h2>Architecture</h2>
         <div class="body-text">
           <h3>MOSS-Audio-Tokenizer</h3>
           <p class="arch-copy">
-            The tokenizer is a causal Transformer codec that compresses
+            The tokenizer is a causal Transformer audio codec that compresses
             <strong>48 kHz stereo</strong> audio into a <strong>12.5 fps</strong>
             RVQ token stream for scalable autoregressive modeling. In the
             report, the encoder and decoder each contain 12 causal
-            Transformer blocks with 10-second sliding-window attention, and
+            Transformer blocks with sliding-window attention, and
             the quantizer uses 16 RVQ layers so the token sequence remains
             compact enough for long-context generation.
           </p>