diff --git a/studio/backend/tests/test_llama_cpp_freshness.py b/studio/backend/tests/test_llama_cpp_freshness.py
index cb17e0d5e7..7f7a433494 100644
--- a/studio/backend/tests/test_llama_cpp_freshness.py
+++ b/studio/backend/tests/test_llama_cpp_freshness.py
@@ -222,6 +222,34 @@ def test_check_prebuilt_freshness_not_stale_when_tag_matches(monkeypatch, tmp_pa
     assert info["latest_tag"] == "b9300"
 
 
+def test_check_prebuilt_freshness_not_stale_when_installed_newer(monkeypatch, tmp_path):
+    """/releases/latest can lag behind out-of-order fork releases; an installed
+    build newer than "latest" must never be flagged stale."""
+    install_dir = tmp_path / "llama.cpp"
+    _write_marker(
+        install_dir,
+        tag = "b9596",
+        installed_at_utc = (datetime.now(tz = timezone.utc) - timedelta(days = 30))
+        .isoformat()
+        .replace("+00:00", "Z"),
+    )
+    bin_path = _fake_binary(install_dir, layout = "root")
+    monkeypatch.setattr(fr, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9594")
+    info = fr.check_prebuilt_freshness(str(bin_path))
+    assert info["stale"] is False
+    assert info["installed_tag"] == "b9596"
+    assert info["latest_tag"] == "b9594"
+
+
+def test_parse_build_number():
+    assert fr.parse_build_number("b9596") == 9596
+    assert fr.parse_build_number(" b9596 ") == 9596
+    assert fr.parse_build_number("9596") is None
+    assert fr.parse_build_number("b9596-custom") is None
+    assert fr.parse_build_number("") is None
+    assert fr.parse_build_number(None) is None
+
+
 def test_check_prebuilt_freshness_not_stale_within_threshold(monkeypatch, tmp_path):
     # Behind by tag but within the 3-day grace window.
     install_dir = tmp_path / "llama.cpp"
diff --git a/studio/backend/tests/test_llama_cpp_update.py b/studio/backend/tests/test_llama_cpp_update.py
index d21d653b76..83ab028d5e 100644
--- a/studio/backend/tests/test_llama_cpp_update.py
+++ b/studio/backend/tests/test_llama_cpp_update.py
@@ -275,6 +275,26 @@ def test_status_up_to_date(monkeypatch, tmp_path):
     assert st["update_available"] is False
 
 
+def test_status_installed_newer_than_latest_release_not_offered(monkeypatch, tmp_path):
+    """GitHub /releases/latest sorts by commit date and can return an older
+    build than the one just installed; never offer that as an "update"."""
+    binary = _write_install(tmp_path, "b9596")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9594")
+    st = upd.get_update_status(force_refresh = True)
+    assert st["installed_tag"] == "b9596"
+    assert st["latest_tag"] == "b9594"
+    assert st["update_available"] is False
+
+
+def test_status_non_numeric_tags_fall_back_to_inequality(monkeypatch, tmp_path):
+    binary = _write_install(tmp_path, "custom-build")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9594")
+    st = upd.get_update_status(force_refresh = True)
+    assert st["update_available"] is True
+
+
 def test_start_update_no_marker_no_prebuilt_refuses(monkeypatch, tmp_path):
     binary = tmp_path / "llama-server"
     binary.write_text("stub")  # no marker
diff --git a/studio/backend/utils/llama_cpp_freshness.py b/studio/backend/utils/llama_cpp_freshness.py
index 3e5066ca2d..556ea71d69 100644
--- a/studio/backend/utils/llama_cpp_freshness.py
+++ b/studio/backend/utils/llama_cpp_freshness.py
@@ -13,6 +13,7 @@ from __future__ import annotations
 
 import json
 import os
+import re
 import time
 from datetime import datetime, timezone
 from pathlib import Path
@@ -43,6 +44,14 @@ def _cache_dir() -> Path:
         return Path.home() / ".unsloth" / "studio" / "cache" / "llama_cpp_freshness"
 
 
+def parse_build_number(tag: object) -> Optional[int]:
+    """Numeric build from a ``bNNNN`` tag, None for anything else."""
+    if not isinstance(tag, str):
+        return None
+    m = re.fullmatch(r"b(\d+)", tag.strip())
+    return int(m.group(1)) if m else None
+
+
 def read_install_marker(binary_path: Optional[str]) -> Optional[dict]:
     """Walk up from binary_path to find UNSLOTH_PREBUILT_INFO.json.
     None = no marker (source build / custom path) or invalid JSON."""
@@ -208,6 +217,14 @@ def check_prebuilt_freshness(
     if not latest or latest == out["installed_tag"]:
         return out
 
+    # GitHub /releases/latest sorts by commit date, not build number, so it can
+    # lag the installed build when releases land out of order; never call a
+    # downgrade stale.
+    installed_build = parse_build_number(out["installed_tag"])
+    latest_build = parse_build_number(latest)
+    if installed_build is not None and latest_build is not None and latest_build <= installed_build:
+        return out
+
     installed_at = _parse_installed_at(out["installed_at_utc"])
     if installed_at is None:
         return out
diff --git a/studio/backend/utils/llama_cpp_update.py b/studio/backend/utils/llama_cpp_update.py
index 55bf4b3d84..58963f0f85 100644
--- a/studio/backend/utils/llama_cpp_update.py
+++ b/studio/backend/utils/llama_cpp_update.py
@@ -10,7 +10,8 @@ the next model load uses it.
 
 Design notes:
 - Detection is delegated to check_prebuilt_freshness(). We surface an
-  ``update_available`` flag (installed_tag != latest_tag) which is laxer than
+  ``update_available`` flag (installed build older than the latest release;
+  tag inequality only when either tag is not bNNNN) which is laxer than
   freshness' ``stale`` (which additionally requires the install to be >= 3 days
   old). The UI shows the "Update llama.cpp" affordance on update_available.
 - The install is slow (download + extract + validate), so it runs on a daemon
@@ -37,6 +38,7 @@ from utils.llama_cpp_freshness import (
     _INSTALL_MARKER_NAME,
     check_prebuilt_freshness,
     latest_published_release,
+    parse_build_number,
     read_install_marker,
     reset_caches,
 )
@@ -286,9 +288,16 @@ def get_update_status(*, force_refresh: bool = False) -> dict:
     freshness = check_prebuilt_freshness(binary)
     installed = freshness.get("installed_tag")
     latest = freshness.get("latest_tag")
-    update_available = bool(
-        freshness.get("has_marker") and installed and latest and installed != latest
-    )
+    # GitHub /releases/latest sorts by commit date, so it can point at an older
+    # build than the one the installer just resolved; compare build numbers and
+    # only offer real upgrades. Tag inequality is the fallback for non-bNNNN tags.
+    installed_build = parse_build_number(installed)
+    latest_build = parse_build_number(latest)
+    if installed_build is not None and latest_build is not None:
+        behind = installed_build < latest_build
+    else:
+        behind = installed != latest
+    update_available = bool(freshness.get("has_marker") and installed and latest and behind)
 
     with _job_lock:
         job = dict(_job)