diff --git a/python/fasttext_module/fasttext/FastText.py b/python/fasttext_module/fasttext/FastText.py
index d4081b4..199d2c4 100644
--- a/python/fasttext_module/fasttext/FastText.py
+++ b/python/fasttext_module/fasttext/FastText.py
@@ -363,6 +363,8 @@ def quantize(
         self.f.quantize(
             input, qout, cutoff, retrain, epoch, lr, thread, verbose, dsub, qnorm
         )
+        # cutoff prunes the dictionary
+        self._words = None
 
     def set_matrices(self, input_matrix, output_matrix):
         """
diff --git a/python/fasttext_module/fasttext/tests/test_nn_cache.py b/python/fasttext_module/fasttext/tests/test_nn_cache.py
new file mode 100644
index 0000000..5fe95f1
--- /dev/null
+++ b/python/fasttext_module/fasttext/tests/test_nn_cache.py
@@ -0,0 +1,64 @@
+# SPDX-FileContributor: Arthit Suriyawongkul
+# SPDX-FileCopyrightText: 2026-present, fasttext-community
+# SPDX-FileType: SOURCE
+# SPDX-License-Identifier: MIT
+
+"""Cached word data must be rebuilt when the model changes."""
+
+import pytest
+
+import fasttext
+import fasttext_pybind
+
+from .helpers import build_supervised_model, get_random_data
+
+
+def _model():
+    # thread=12: thread <= 10 leaves the input matrix partly uninitialized.
+    data = get_random_data(3000, max_vocab_size=600)
+    return build_supervised_model(data, {"thread": 12, "dim": 16, "verbose": 0})
+
+
+def _filled_model():
+    model = _model()
+    model.get_nearest_neighbors(model.words[1])  # fills the caches
+    return model
+
+
+def _reload(model, tmp_path):
+    path = str(tmp_path / "model.bin")
+    model.save_model(path)
+    return fasttext.load_model(path)
+
+
+def _assert_same_nn(model, expected):
+    word = expected.words[1]  # words[0] is the end-of-sentence token
+    assert model.get_nearest_neighbors(word) == expected.get_nearest_neighbors(word)
+
+
+# cutoff=300 also prunes the dictionary (quantize needs >= 256 rows).
+@pytest.mark.parametrize("cutoff", [0, 300])
+def test_quantize_resets_caches(tmp_path, cutoff):
+    model = _filled_model()
+    model.quantize(cutoff=cutoff)
+    expected = _reload(model, tmp_path)
+    assert model.words == expected.words
+    _assert_same_nn(model, expected)
+
+
+def test_load_model_resets_cache(tmp_path):
+    other = _reload(_model(), tmp_path)
+    model = _filled_model()
+    model.f.loadModel(str(tmp_path / "model.bin"))
+    _assert_same_nn(model, other)
+
+
+def test_train_resets_cache(tmp_path):
+    model = _filled_model()
+    train_txt = tmp_path / "train.txt"
+    data = get_random_data(3000)
+    train_txt.write_text("".join(f"__label__{line}\n" for line in data))
+    args = model.f.getArgs()
+    args.input = str(train_txt)
+    fasttext_pybind.train(model.f, args)
+    _assert_same_nn(model, _reload(model, tmp_path))
diff --git a/src/fasttext.cc b/src/fasttext.cc
index a659c55..dfad387 100644
--- a/src/fasttext.cc
+++ b/src/fasttext.cc
@@ -293,6 +293,7 @@ namespace fasttext
 
   void FastText::loadModel(std::istream &in)
   {
+    wordVectors_.reset();
     args_ = std::make_shared<Args>();
     input_ = std::make_shared<DenseMatrix>();
     output_ = std::make_shared<DenseMatrix>();
@@ -400,6 +401,10 @@ namespace fasttext
         std::dynamic_pointer_cast<DenseMatrix>(output_);
     bool normalizeGradient = (args_->model == model_name::sup);
 
+    // input_ is replaced below (and dict_ pruned if cutoff > 0), so the
+    // cached word vectors used by getNN/getAnalogies are stale either way.
+    wordVectors_.reset();
+
     if (qargs.cutoff > 0 && static_cast<size_t>(qargs.cutoff) < static_cast<size_t>(input->size(0)))
     {
       auto idx = selectEmbeddings(qargs.cutoff);
@@ -892,6 +897,7 @@ namespace fasttext
           "bucket must be > 0 when using subwords (maxn > 0) "
           "or word n-grams (wordNgrams > 1)");
     }
+    wordVectors_.reset();
     args_ = std::make_shared<Args>(args);
     dict_ = std::make_shared<Dictionary>(args_);
     if (args_->input == "-")