diff --git a/isovar/__init__.py b/isovar/__init__.py
index c034439..0257a83 100644
--- a/isovar/__init__.py
+++ b/isovar/__init__.py
@@ -10,7 +10,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "1.4.23"
+__version__ = "1.4.24"
 
 
 from .allele_read import AlleleRead
diff --git a/isovar/assembly.py b/isovar/assembly.py
index 3373315..ed2fc52 100644
--- a/isovar/assembly.py
+++ b/isovar/assembly.py
@@ -37,6 +37,62 @@
 
 logger = get_logger(__name__)
 
+DEFAULT_ASSEMBLY_KMER_SIZE = 15
+
+
+def _sequence_kmers(sequence, k):
+    return {
+        sequence[i:i + k]
+        for i in range(len(sequence) - k + 1)
+    }
+
+
+def _greedy_merge_candidate_pairs(
+        variant_sequences,
+        min_overlap_size=MIN_VARIANT_SEQUENCE_ASSEMBLY_OVERLAP_SIZE,
+        max_kmer_size=DEFAULT_ASSEMBLY_KMER_SIZE):
+    """
+    Use an exact k-mer index to skip obviously impossible pairwise comparisons.
+
+    Any truly mergeable pair must share at least one exact k-mer whose length
+    is no greater than the overlap threshold and no greater than the shortest
+    eligible sequence. Final merge eligibility is still decided by
+    VariantSequence.combine.
+    """
+    if len(variant_sequences) < 2:
+        return []
+
+    eligible_lengths = [
+        len(variant_sequence)
+        for variant_sequence in variant_sequences
+        if len(variant_sequence) >= min_overlap_size
+    ]
+    if not eligible_lengths:
+        return []
+
+    k = max(1, min(max_kmer_size, min_overlap_size, min(eligible_lengths)))
+    kmer_to_indices = defaultdict(set)
+    sequence_kmers = []
+
+    for i, variant_sequence in enumerate(variant_sequences):
+        if len(variant_sequence) < min_overlap_size:
+            kmers = ()
+        else:
+            kmers = tuple(_sequence_kmers(variant_sequence.sequence, k))
+            for kmer in kmers:
+                kmer_to_indices[kmer].add(i)
+        sequence_kmers.append(kmers)
+
+    candidate_pairs = set()
+    for i, kmers in enumerate(sequence_kmers):
+        candidate_indices = set()
+        for kmer in kmers:
+            candidate_indices.update(kmer_to_indices[kmer])
+        for j in candidate_indices:
+            if j > i:
+                candidate_pairs.add((i, j))
+    return sorted(candidate_pairs)
+
 
 def greedy_merge_helper(
         variant_sequences,
@@ -51,16 +107,17 @@ def greedy_merge_helper(
     were successfully merged.
     """
     candidates = []
-    for i in range(len(variant_sequences)):
-        for j in range(i + 1, len(variant_sequences)):
-            combined = variant_sequences[i].combine(
-                variant_sequences[j], min_overlap_size=min_overlap_size)
-            if combined is not None:
-                overlap = (
-                    len(variant_sequences[i])
-                    + len(variant_sequences[j])
-                    - len(combined))
-                candidates.append((overlap, len(combined.reads), i, j, combined))
+    for i, j in _greedy_merge_candidate_pairs(
+            variant_sequences,
+            min_overlap_size=min_overlap_size):
+        combined = variant_sequences[i].combine(
+            variant_sequences[j], min_overlap_size=min_overlap_size)
+        if combined is not None:
+            overlap = (
+                len(variant_sequences[i])
+                + len(variant_sequences[j])
+                - len(combined))
+            candidates.append((overlap, len(combined.reads), i, j, combined))
 
     if not candidates:
         return list(variant_sequences), False
diff --git a/tests/test_assembly.py b/tests/test_assembly.py
index 71005f7..f8381cb 100644
--- a/tests/test_assembly.py
+++ b/tests/test_assembly.py
@@ -20,6 +20,7 @@
 from isovar.assembly import (
     iterative_overlap_assembly,
     greedy_merge,
+    greedy_merge_helper,
     collapse_substrings
 )
 
@@ -365,6 +366,36 @@ def test_greedy_merge_deterministic_across_input_orders():
                 trial, reference_all_reads, result_all_reads)
 
 
+def test_greedy_merge_helper_skips_impossible_pairs(monkeypatch):
+    variant_sequences = [
+        VariantSequence(prefix="AAAAAA", alt="X", suffix="AAAAAA", reads={"a"}),
+        VariantSequence(prefix="CCCCCC", alt="X", suffix="CCCCCC", reads={"c"}),
+        VariantSequence(prefix="GGGGGG", alt="X", suffix="GGGGGG", reads={"g"}),
+        VariantSequence(prefix="TTTTTT", alt="X", suffix="TTTTTT", reads={"t"}),
+    ]
+
+    combine_call_count = 0
+    original_combine = VariantSequence.combine
+
+    def counting_combine(self, other_sequence, min_overlap_size=1):
+        nonlocal combine_call_count
+        combine_call_count += 1
+        return original_combine(
+            self,
+            other_sequence,
+            min_overlap_size=min_overlap_size)
+
+    monkeypatch.setattr(VariantSequence, "combine", counting_combine)
+
+    result, merged_any = greedy_merge_helper(
+        variant_sequences,
+        min_overlap_size=3)
+
+    eq_(combine_call_count, 0)
+    eq_(merged_any, False)
+    eq_(result, variant_sequences)
+
+
 def test_greedy_merge_prefers_larger_overlap():
     """
     When a sequence could merge with two partners, the larger overlap