Re-convert with fixed rotary q/k permutation (baseRT-internal#114); NIAH-validated at 2k/8k. Q4 now uses the default-q4 quality profile (embeddings/head at higher precision).

Llama-3.2-3B-Instruct-Q4.base

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a0d5be46b5e13ce488e351cb8d769d3402250ee8adf3ff1fa41e70328f1fac8a
-size 1814222848
+oid sha256:20c803b2eeee70a7dada67ca0ee2d75081d95d0e35bac27d1857f9e3fdd84615
+size 2380601344