diff --git a/review-enrichment/src/analyzers/asset-weight.ts b/review-enrichment/src/analyzers/asset-weight.ts
index 190a5fbe07..f142a3ebea 100644
--- a/review-enrichment/src/analyzers/asset-weight.ts
+++ b/review-enrichment/src/analyzers/asset-weight.ts
@@ -72,6 +72,14 @@ const BINARY_EXTS = new Set([
   "node",
   "jar",
   "class",
+  // Serialized ML model / checkpoint weight formats — routinely hundreds of MB to multi-GB, the heaviest
+  // binary blobs a PR can commit, and their bytes never appear in the textual diff.
+  "safetensors",
+  "gguf",
+  "onnx",
+  "pt",
+  "pth",
+  "ckpt",
 ]);
 
 interface ScanOptions {
diff --git a/review-enrichment/test/asset-weight.test.ts b/review-enrichment/test/asset-weight.test.ts
index 620d14d4e2..01f9b6a069 100644
--- a/review-enrichment/test/asset-weight.test.ts
+++ b/review-enrichment/test/asset-weight.test.ts
@@ -28,6 +28,19 @@ test("isBinaryAsset flags genuine binary extensions and ignores text/case", () =
   assert.equal(isBinaryAsset("cache/model.zst"), true);
   assert.equal(isBinaryAsset("dist/bundle.tar.zst"), true);
   assert.equal(isBinaryAsset("cache/model.ZST"), true);
+  // Serialized ML model / checkpoint weight formats are heavy binary blobs (siblings of the other binaries),
+  // and the match stays case-insensitive.
+  for (const p of [
+    "models/llama.safetensors",
+    "models/llama.gguf",
+    "models/resnet.onnx",
+    "models/model.pt",
+    "checkpoints/epoch10.pth",
+    "checkpoints/state.ckpt",
+    "models/LLAMA.SAFETENSORS",
+  ]) {
+    assert.equal(isBinaryAsset(p), true, p);
+  }
   // Extension match is case-insensitive.
   assert.equal(isBinaryAsset("assets/HERO.PNG"), true);
   // Text formats whose bytes are already in the diff are NOT binary assets.