diff --git a/gpu/modal_app.py b/gpu/modal_app.py
index 231e50a..36b09af 100644
--- a/gpu/modal_app.py
+++ b/gpu/modal_app.py
@@ -148,14 +148,22 @@ def cache_msa(label: str, protein_seq: str, ligand_smiles: str, cofactor_ccds: l
     out = work / "out"
     subprocess.run(["boltz", "predict", str(work / "in.yaml"), "--use_msa_server",
                     "--cache", boltz_cache, "--out_dir", str(out), "--output_format", "pdb"], check=True)
-    a3m = next(out.rglob("*.a3m"), None)
-    if a3m is None:
+    # Pick the largest a3m that actually looks like FASTA/a3m text (Boltz writes several files;
+    # some are binary and crash its INPUT parser with KeyError '\x00'). Strip null bytes too.
+    candidates = sorted(out.rglob("*.a3m"), key=lambda p: p.stat().st_size, reverse=True)
+    chosen = None
+    for c in candidates:
+        raw = c.read_bytes()
+        if raw[:1] == b">" or b"\n>" in raw[:512]:
+            chosen = raw.replace(b"\x00", b"")
+            break
+    if chosen is None:
         weights.commit()
-        raise RuntimeError("MSA generation produced no .a3m")
+        raise RuntimeError(f"no valid a3m; candidates={[(str(p), p.stat().st_size) for p in candidates]}")
     msa_dir = Path(WEIGHTS) / "msa"
     msa_dir.mkdir(parents=True, exist_ok=True)
     dest = msa_dir / f"{label}.a3m"
-    shutil.copy(a3m, dest)
+    dest.write_bytes(chosen)
     weights.commit()
     return str(dest)