Update: you were right to suggest checking the hash. My cached GGUF blob was corrupt. HF expected SHA256: 9188a71055550f1e60b875d02b7abb63625ac11b4a6f148d6b22b3b28ba3d335 My old local blob hashed to: 20e9ffda0c1a0fb5b6ed9cc445834e5c3e98a1f9ffe4a64edf319cbd0aa85fba I moved the blob aside, force-redownloaded with hf download --force-download , and rebuilt latest llama.cpp master after the Gemma 4 MTP merge. Result: main 31B QAT GGUF now works. No more repeated <unused49> . Te...