diff --git a/fastembed/common/onnx_model.py b/fastembed/common/onnx_model.py
index d357f2c1..3f99a210 100644
--- a/fastembed/common/onnx_model.py
+++ b/fastembed/common/onnx_model.py
@@ -156,6 +156,19 @@ def add_extra_session_options(
         if "enable_cpu_mem_arena" in extra_options:
             session_options.enable_cpu_mem_arena = extra_options["enable_cpu_mem_arena"]
 
+    def _check_output_finite(self, output: NumpyArray) -> None:
+        """Validate that the model output contains only finite numbers (no NaN or Inf).
+
+        Raises:
+            RuntimeError: If non-finite values (NaN/Inf) are detected in the output.
+        """
+        if not np.all(np.isfinite(output)):
+            providers = self.model.get_providers() if self.model is not None else None
+            raise RuntimeError(
+                f"Model '{self.model_name}' produced non-finite (NaN/Inf) embeddings "
+                f"with provider(s) '{providers}'."
+            )
+
     def load_onnx_model(self) -> None:
         raise NotImplementedError("Subclasses must implement this method")
 
diff --git a/fastembed/text/builtin_sentence_embedding.py b/fastembed/text/builtin_sentence_embedding.py
index 430ea7ca..bcf83370 100644
--- a/fastembed/text/builtin_sentence_embedding.py
+++ b/fastembed/text/builtin_sentence_embedding.py
@@ -51,7 +51,9 @@ def _post_process_onnx_output(
     def _run_model(
         self, onnx_input: dict[str, Any], onnx_output_names: list[str] | None = None
     ) -> NumpyArray:
-        return self.model.run(onnx_output_names, onnx_input)[1]  # type: ignore[union-attr]
+        model_output = self.model.run(onnx_output_names, onnx_input)[1]  # type: ignore[union-attr]
+        self._check_output_finite(model_output)
+        return model_output
 
 
 class BuiltinSentenceEmbeddingWorker(OnnxTextEmbeddingWorker):
diff --git a/fastembed/text/onnx_text_model.py b/fastembed/text/onnx_text_model.py
index 10a4aa17..cf6b11f7 100644
--- a/fastembed/text/onnx_text_model.py
+++ b/fastembed/text/onnx_text_model.py
@@ -105,7 +105,9 @@ def onnx_embed(
     def _run_model(
         self, onnx_input: dict[str, Any], onnx_output_names: list[str] | None = None
     ) -> NumpyArray:
-        return self.model.run(onnx_output_names, onnx_input)[0]  # type: ignore[union-attr]
+        model_output = self.model.run(onnx_output_names, onnx_input)[0]  # type: ignore[union-attr]
+        self._check_output_finite(model_output)
+        return model_output
 
     def _embed_documents(
         self,
diff --git a/tests/test_text_onnx_embeddings.py b/tests/test_text_onnx_embeddings.py
index 8744b617..dad79acb 100644
--- a/tests/test_text_onnx_embeddings.py
+++ b/tests/test_text_onnx_embeddings.py
@@ -308,3 +308,71 @@ def test_token_count(model_cache, model_name) -> None:
         doc_token_count = model.token_count(documents)
         assert first_doc_token_count + second_doc_token_count == doc_token_count
         assert doc_token_count == model.token_count(documents, batch_size=1)
+
+
+# ---------------------------------------------------------------------------
+# Regression tests for issue #688 - non-finite embedding guard
+# ---------------------------------------------------------------------------
+# All four tests use monkeypatch to replace the underlying ort.InferenceSession.run
+# so they are deterministic, fast, and independent of any real model download or
+# specific ONNX Runtime version.
+# ---------------------------------------------------------------------------
+
+_BGE_SMALL = "BAAI/bge-small-en-v1.5"
+
+
+def _load_bge_small() -> TextEmbedding:
+    """Load a small, widely-cached model. Uses lazy_load=False to ensure the
+    ONNX session is initialised before monkeypatching."""
+    return TextEmbedding(model_name=_BGE_SMALL, lazy_load=False, cuda=False)
+
+
+def test_non_finite_nan_raises(monkeypatch) -> None:
+    """NaN values must raise RuntimeError with model name and provider (issue #688).
+
+    Verifies the full diagnostic message so users can immediately identify which
+    model/provider combination produced the bad output.
+    """
+    model = _load_bge_small()
+    monkeypatch.setattr(
+        model.model.model,
+        "run",
+        lambda *args, **kwargs: [np.full((1, 5, 384), np.nan, dtype=np.float32)],
+    )
+    with pytest.raises(
+        RuntimeError,
+        match=r"Model 'BAAI/bge-small-en-v1\.5' produced non-finite.*CPUExecutionProvider",
+    ):
+        list(model.embed(["test"]))
+
+
+def test_non_finite_pos_inf_raises(monkeypatch) -> None:
+    """+Inf values in ONNX output must raise RuntimeError (issue #688)."""
+    model = _load_bge_small()
+    monkeypatch.setattr(
+        model.model.model,
+        "run",
+        lambda *args, **kwargs: [np.full((1, 5, 384), np.inf, dtype=np.float32)],
+    )
+    with pytest.raises(RuntimeError, match="non-finite"):
+        list(model.embed(["test"]))
+
+
+def test_non_finite_neg_inf_raises(monkeypatch) -> None:
+    """-Inf values in ONNX output must raise RuntimeError (issue #688)."""
+    model = _load_bge_small()
+    monkeypatch.setattr(
+        model.model.model,
+        "run",
+        lambda *args, **kwargs: [np.full((1, 5, 384), -np.inf, dtype=np.float32)],
+    )
+    with pytest.raises(RuntimeError, match="non-finite"):
+        list(model.embed(["test"]))
+
+
+def test_non_finite_finite_output_succeeds() -> None:
+    """Sanity check: finite ONNX output must not raise (issue #688)."""
+    model = _load_bge_small()
+    results = list(model.embed(["test"]))
+    assert len(results) == 1
+    assert np.all(np.isfinite(results[0]))