diff --git a/tests/test_generate_kwarg_gate.py b/tests/test_generate_kwarg_gate.py
index 6d1379d3a9..242fcebfbf 100644
--- a/tests/test_generate_kwarg_gate.py
+++ b/tests/test_generate_kwarg_gate.py
@@ -130,6 +130,50 @@ def test_generate_kwarg_gate():
         assert got is expected, f"{name}: got {got}, expected {expected}"
 
 
+# --- v5 logits-to-keep filtering ------------------------------------------
+# transformers >= 5 injects logits_to_keep=1 in generate() itself, but the
+# injection is guarded by `"logits_to_keep" not in model_kwargs`, so it is a
+# DEFAULT. An explicit caller value must survive: popping unconditionally turns
+# logits_to_keep=0 (give me the full sequence) into 1 without telling anyone.
+# The only values that must be stripped are the ones the strict validator would
+# raise on, which is exactly what the gate above predicts.
+
+def _filter_logits_kwargs(model, kwargs):
+    """The v5 branch of unsloth_base_fast_generate, as a testable function."""
+    for key in ("logits_to_keep", "num_logits_to_keep"):
+        if key in kwargs and not accepts(model, key):
+            kwargs.pop(key, None)
+    return kwargs
+
+
+def test_v5_preserves_a_supported_caller_value():
+    model = PrepHasKwargs_ForwardHasKey()
+    # 0 means "all logits"; silently rewriting it to 1 changes the output shape.
+    assert _filter_logits_kwargs(model, {"logits_to_keep": 0}) == {"logits_to_keep": 0}
+    assert _filter_logits_kwargs(model, {"logits_to_keep": 5}) == {"logits_to_keep": 5}
+
+
+def test_v5_strips_a_value_the_model_would_reject():
+    # num_logits_to_keep was renamed away in v5, so the validator raises on it.
+    model = PrepHasKwargs_ForwardHasKey()
+    assert _filter_logits_kwargs(model, {"num_logits_to_keep": 1}) == {}
+    # A VLM whose top-level forward has no logits_to_keep at all.
+    assert _filter_logits_kwargs(NoPrepare(), {"logits_to_keep": 1}) == {}
+
+
+def test_v5_leaves_other_kwargs_alone():
+    model = PrepHasKwargs_ForwardHasKey()
+    out = _filter_logits_kwargs(model, {"logits_to_keep": 2, "max_new_tokens": 8})
+    assert out == {"logits_to_keep": 2, "max_new_tokens": 8}
+
+
+def test_source_has_no_unconditional_pop():
+    src = open(VISION).read()
+    assert 'kwargs.pop("logits_to_keep", None)\n        kwargs.pop("num_logits_to_keep", None)' not in src, (
+        "the v5 branch must not drop caller-supplied logits_to_keep unconditionally"
+    )
+
+
 if __name__ == "__main__":
     test_generate_kwarg_gate()
     for name, _, _, _ in CASES:
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index f20f32cc3e..effe633302 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -399,8 +399,8 @@ def unsloth_base_fast_generate(self, *args, **kwargs):
         kwargs.pop("mm_token_type_ids", None)
 
     # VLMs do not allow logits_to_keep. transformers >= 5.0 sets it itself in
-    # generate() after _validate_model_kwargs, so pre-injecting makes the strict
-    # validator raise on PEFT models. Skip on v5+ and strip any leaked kwarg.
+    # generate(), so pre-injecting is redundant there, and the arch walk below
+    # can pick a key the top-level model rejects. Skip the injection on v5+.
     if Version(transformers_version) < Version("5.0.0.dev0"):
         global NUM_LOGITS_TO_KEEP
         if arch not in NUM_LOGITS_TO_KEEP:
@@ -422,8 +422,18 @@ def unsloth_base_fast_generate(self, *args, **kwargs):
         if key is not None and key not in kwargs and _unsloth_generate_accepts_kwarg(self, key):
             kwargs[key] = 1
     else:
-        kwargs.pop("logits_to_keep", None)
-        kwargs.pop("num_logits_to_keep", None)
+        # v5's own injection (generation/utils.py) is guarded by
+        # `"logits_to_keep" not in model_kwargs`, so it is a default, not an
+        # override: an explicit caller value survives and must not be dropped.
+        # Popping unconditionally silently rewrites logits_to_keep=0 (full
+        # sequence) into 1. Only strip a key this model would reject, which is
+        # what the strict validator raises on: num_logits_to_keep everywhere
+        # (renamed away in v5), and logits_to_keep on the VLMs whose top-level
+        # forward does not take it.
+        for _logits_kwarg in ("logits_to_keep", "num_logits_to_keep"):
+            if _logits_kwarg in kwargs and \
+                    not _unsloth_generate_accepts_kwarg(self, _logits_kwarg):
+                kwargs.pop(_logits_kwarg, None)
 
     model_eos_token_id = getattr(self.config, "eos_token_id", None)
     if model_eos_token_id is not None and hasattr(model_eos_token_id, "__iter__"):