diff --git a/src/transformers/generation/utils.py b/src/transformers/generation/utils.py
index 95cee542b291..cc88ca502dd8 100644
--- a/src/transformers/generation/utils.py
+++ b/src/transformers/generation/utils.py
@@ -1788,6 +1788,12 @@ def _prepare_generation_config(
         model_kwargs.update({"output_attentions": output_attentions} if output_attentions else {})
         model_kwargs.update({"output_hidden_states": output_hidden_states} if output_hidden_states else {})
 
+        # Enforce deterministic greedy decoding if do_sample=False and num_beams = 1
+        if generation_config.do_sample is False and generation_config.num_beams == 1:
+            generation_config.temperature = 1.0
+            generation_config.top_k = 0
+            generation_config.top_p = 1.0
+
         return generation_config, model_kwargs
 
     def _get_initial_cache_position(self, seq_length, device, model_kwargs):
diff --git a/tests/models/qwen3/test_modeling_qwen3.py b/tests/models/qwen3/test_modeling_qwen3.py
index ba937656d3a6..05c5e746ba47 100644
--- a/tests/models/qwen3/test_modeling_qwen3.py
+++ b/tests/models/qwen3/test_modeling_qwen3.py
@@ -718,3 +718,20 @@ def test_600m_generation(self):
         new_generated_ids = model.generate(input_ids, max_new_tokens=50)[:, input_ids.shape[1] :]
         with self.subTest("Eager matches flash attention"):
             torch.testing.assert_close(generated_ids, new_generated_ids, rtol=1e-4, atol=1e-4)
+
+    def test_qwen3_greedy_determinism(self):
+        """
+        Ensures Qwen3 generate is deterministic when do_sample=False (greedy decoding as per HFs documentation).
+        """
+        tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B-Base", use_fast=False)
+        model = Qwen3ForCausalLM.from_pretrained("Qwen/Qwen3-0.6B-Base", device_map="auto")
+        inputs = tokenizer("hello", return_tensors="pt")
+
+        cfg = GenerationConfig(do_sample=False, num_beams=1, max_new_tokens=20)
+
+        out1 = model.generate(**inputs, generation_config=cfg)
+        out2 = model.generate(**inputs, generation_config=cfg)
+
+        assert torch.equal(out1, out2), (
+            "Qwen3 should produce deterministic outputs with do_sample=False and num_beams=1"
+        )