quantization 2 Turbo3 + MTP: Merging Two llama.cpp Forks May 15, 2026 Nemotron 30B on 24 GB: Benchmarks and a Quantization Quirk Apr 29, 2026