diff --git a/gui/chatblockmodel.cpp b/gui/chatblockmodel.cpp
index 68372bb..126abb4 100644
--- a/gui/chatblockmodel.cpp
+++ b/gui/chatblockmodel.cpp
@@ -107,17 +107,16 @@ void ChatBlockModel::setRawText(const QString &text)
 {
     if (text == m_rawText) return;
 
-    // If text is just appended (common during streaming), only reparse the last block
+    // If text is just appended (common during streaming), only reparse from the last block
     bool isAppend = text.startsWith(m_rawText) && !m_rawText.isEmpty();
     m_rawText = text;
 
     if (isAppend && !m_allBlocks.isEmpty()) {
-        // Re-parse only: find where the last block started in the text and reparse from there
-        // Simpler: just reparse the full text but it's cached via incremental diff
-        reparse();
+        reparseIncremental();
     } else {
         // Full reparse (session switch, initial load)
         m_allBlocks.clear();
+        m_lastBlockTextOffset = 0;
         beginResetModel();
         m_blocks.clear();
         m_visibleStart = 0;
@@ -130,10 +129,58 @@ void ChatBlockModel::setRawText(const QString &text)
 static constexpr int INITIAL_BLOCKS = 20;
 static constexpr int LOAD_MORE_COUNT = 30;
 
+void ChatBlockModel::reparseIncremental()
+{
+    // Only re-parse from where the last block started — O(tail) not O(full text)
+    QString tail = m_rawText.mid(m_lastBlockTextOffset);
+    QList<ChatBlock> tailBlocks = parseBlocks(tail);
+
+    // Replace the last block (which may have grown) and append any new blocks
+    if (!m_allBlocks.isEmpty())
+        m_allBlocks.removeLast();
+    m_allBlocks.append(tailBlocks);
+
+    // Update offset: last block starts at the last header in the tail
+    if (tailBlocks.size() > 1) {
+        // Multiple blocks parsed from tail — advance offset to where the last one starts
+        // Find the last header boundary in m_rawText
+        static const QRegularExpression headerRe(
+            R"(^\[(user|assistant|reasoning|call|tool|error|info)(?::([^\]]*))\])",
+            QRegularExpression::MultilineOption);
+        auto it = headerRe.globalMatch(m_rawText);
+        int lastPos = 0;
+        while (it.hasNext()) {
+            auto match = it.next();
+            lastPos = match.capturedStart();
+        }
+        m_lastBlockTextOffset = lastPos;
+    }
+
+    updateVisibleWindow();
+}
+
 void ChatBlockModel::reparse()
 {
     m_allBlocks = parseBlocks(m_rawText);
+    // Track where the last block starts for future incremental parses
+    if (!m_allBlocks.isEmpty()) {
+        static const QRegularExpression headerRe(
+            R"(^\[(user|assistant|reasoning|call|tool|error|info)(?::([^\]]*))\])",
+            QRegularExpression::MultilineOption);
+        auto it = headerRe.globalMatch(m_rawText);
+        int lastPos = 0;
+        while (it.hasNext()) {
+            auto match = it.next();
+            lastPos = match.capturedStart();
+        }
+        m_lastBlockTextOffset = lastPos;
+    }
 
+    updateVisibleWindow();
+}
+
+void ChatBlockModel::updateVisibleWindow()
+{
     // Determine visible window: last INITIAL_BLOCKS (or expand if already showing more)
     int desiredStart;
     if (m_blocks.isEmpty()) {
diff --git a/gui/chatblockmodel.h b/gui/chatblockmodel.h
index 73ba5ff..6f3db54 100644
--- a/gui/chatblockmodel.h
+++ b/gui/chatblockmodel.h
@@ -74,6 +74,8 @@ signals:
 
 private:
     void reparse();
+    void reparseIncremental();
+    void updateVisibleWindow();
     QList<ChatBlock> parseBlocks(const QString &text) const;
     QVariantList parseSubBlocks(const QString &content) const;
 
@@ -81,6 +83,7 @@ private:
     QList<ChatBlock> m_allBlocks;   // all parsed blocks
     QList<ChatBlock> m_blocks;      // visible window (tail of m_allBlocks)
     int m_visibleStart = 0;         // index into m_allBlocks where visible window starts
+    int m_lastBlockTextOffset = 0;  // char offset in m_rawText where last block starts
     mutable QHash<QString, QString> m_highlightCache;
 };