diff --git a/static/messages.js b/static/messages.js
index 284ca5b..0209c2f 100644
--- a/static/messages.js
+++ b/static/messages.js
@@ -84,6 +84,11 @@ async function send(){
   let assistantText='';
   let assistantRow=null;
   let assistantBody=null;
+  // Thinking tag patterns for streaming display
+  const _thinkPairs=[
+    {open:'<think>',close:'</think>'},
+    {open:'<|channel>thought\n',close:'<channel|>'}
+  ];
 
   function ensureAssistantRow(){
     if(assistantRow)return;
@@ -106,12 +111,36 @@ async function send(){
 
   // rAF-throttled rendering: buffer tokens, render at most once per frame
   let _renderPending=false;
+  // Extract display text from assistantText, stripping completed thinking blocks
+  // and hiding content still inside an open thinking block.
+  function _streamDisplay(){
+    const raw=assistantText;
+    for(const {open,close} of _thinkPairs){
+      if(raw.startsWith(open)){
+        const ci=raw.indexOf(close,open.length);
+        if(ci!==-1){
+          // Thinking block complete — strip it, show the rest
+          return raw.slice(ci+close.length).replace(/^\s+/,'');
+        }
+        // Still inside thinking block — show placeholder
+        return '';
+      }
+      // Hide partial tag prefixes while streaming so users don't see
+      // `<thi`, `<think`, etc. before the model finishes the token.
+      if(open.startsWith(raw)) return '';
+    }
+    return raw;
+  }
   function _scheduleRender(){
     if(_renderPending) return;
     _renderPending=true;
     requestAnimationFrame(()=>{
       _renderPending=false;
-      if(assistantBody) assistantBody.innerHTML=renderMd(assistantText);
+      if(assistantBody){
+        const txt=_streamDisplay();
+        const isThinking=!txt&&assistantText.length>0;
+        assistantBody.innerHTML=txt?renderMd(txt):(isThinking?'<span style="color:var(--muted);font-size:13px">Thinking\u2026</span>':'');
+      }
       scrollIfPinned();
     });
   }
diff --git a/static/ui.js b/static/ui.js
index f318a32..3b5ab78 100644
--- a/static/ui.js
+++ b/static/ui.js
@@ -503,6 +503,22 @@ function renderMessages(){
     if(!thinkingText && m.reasoning){
       thinkingText=m.reasoning;
     }
+    // Parse inline thinking tags from plain text: <think>...</think> (DeepSeek, QwQ, etc.)
+    // and Gemma 4 channel tokens: <|channel>thought\n...<channel|>
+    if(!thinkingText && typeof content==='string'){
+      const thinkMatch=content.match(/^<think>([\s\S]*?)<\/think>\s*/);
+      if(thinkMatch){
+        thinkingText=thinkMatch[1].trim();
+        content=content.slice(thinkMatch[0].length);
+      }
+      if(!thinkingText){
+        const gemmaMatch=content.match(/^<\|channel>thought\n([\s\S]*?)<channel\|>\s*/);
+        if(gemmaMatch){
+          thinkingText=gemmaMatch[1].trim();
+          content=content.slice(gemmaMatch[0].length);
+        }
+      }
+    }
     const isUser=m.role==='user';
     const isLastAssistant=!isUser&&vi===visWithIdx.length-1;
     // Render thinking card before the assistant message (collapsed by default)