diff --git a/simpler_setup/toolchain.py b/simpler_setup/toolchain.py
index e887b4ce7f..9d2eb25cbb 100644
--- a/simpler_setup/toolchain.py
+++ b/simpler_setup/toolchain.py
@@ -167,7 +167,7 @@ def get_compile_flags(self, core_type: str = "aiv", **kwargs) -> list[str]:
         else:
             raise ValueError(f"Unknown platform: {self.platform}. Supported: a2a3, a2a3sim, a5, a5sim")
 
-        return [
+        flags = [
             "-c",
             "-O3",
             "-g",
@@ -189,6 +189,17 @@ def get_compile_flags(self, core_type: str = "aiv", **kwargs) -> list[str]:
             "-cce-aicore-dcci-insert-for-scalar=false",
             "-DMEMORY_BASE",
         ]
+        # A5: VF fusion can reorder same-V ops across expand/abs/sinkhorn and
+        # break DeepSeek-V4 correctness; keep it off by default for a5/a5sim.
+        #
+        # Scope: per-kernel incore compilation only. Do NOT mirror this into the
+        # platform AICore build (src/a5/platform/onboard/aicore/CMakeLists.txt) --
+        # its only VF construct is the __simd_vf__ meta anchor in simt_anchor.h,
+        # whose fusion behaviour is what makes bisheng tag the entry
+        # SIMD_SIMT_MIX_VF(4); disabling fusion there breaks that classification.
+        if self.platform in ("a5", "a5sim"):
+            flags.append("--cce-simd-vf-fusion=false")
+        return flags
 
     def get_cmake_args(self) -> list[str]:
         return [
diff --git a/tests/ut/py/test_toolchain.py b/tests/ut/py/test_toolchain.py
index 1bd9a432b1..b79af7bbd5 100644
--- a/tests/ut/py/test_toolchain.py
+++ b/tests/ut/py/test_toolchain.py
@@ -185,3 +185,41 @@ def test_path_with_spaces_is_quoted(self, monkeypatch):
         flags_arg = next(a for a in args if a.startswith("-DCMAKE_C_FLAGS="))
         # shlex.join must re-quote the path so CMake re-parses it as one token.
         assert "'/opt/my compilers/compat'" in flags_arg
+
+
+class TestCCECToolchainCompileFlags:
+    """A5 must disable SIMD VF fusion; A2/A3 must be left untouched.
+
+    VF fusion can reorder same-V vector ops (expand/abs/sinkhorn) and silently
+    break DeepSeek-V4 numerics on Ascend950. The failure mode is a wrong result
+    rather than a compile error, so guard the flag explicitly.
+    """
+
+    @pytest.fixture
+    def ccec(self, monkeypatch):
+        """Build a CCECToolchain without needing a real CANN install."""
+
+        def _make(platform):
+            # Constructor probes ASCEND_HOME_PATH and the ccec/ld.lld binaries.
+            monkeypatch.setattr(
+                "simpler_setup.toolchain.env_manager.get", lambda _k: "/nonexistent/ascend"
+            )
+            monkeypatch.setattr("simpler_setup.toolchain.os.path.isfile", lambda _p: True)
+            from simpler_setup.toolchain import CCECToolchain  # noqa: PLC0415
+
+            return CCECToolchain(platform)
+
+        return _make
+
+    @pytest.mark.parametrize("platform", ["a5", "a5sim"])
+    def test_a5_disables_vf_fusion(self, ccec, platform):
+        assert "--cce-simd-vf-fusion=false" in ccec(platform).get_compile_flags()
+
+    @pytest.mark.parametrize("platform", ["a2a3", "a2a3sim"])
+    def test_a2a3_keeps_vf_fusion(self, ccec, platform):
+        assert "--cce-simd-vf-fusion=false" not in ccec(platform).get_compile_flags()
+
+    def test_a5_flag_applies_to_both_core_types(self, ccec):
+        toolchain = ccec("a5")
+        for core_type in ("aiv", "aic"):
+            assert "--cce-simd-vf-fusion=false" in toolchain.get_compile_flags(core_type)