Puzzle-75B, NVIDIA's Mamba/MoE/attention hybrid, now runs in llama.cpp
NVIDIA's Puzzle-75B-A9B, a 75B model mixing Mamba, sparse MoE and attention layers, is now loadable in llama.cpp via PR #25444. Community GGUF quants built on that branch are already live, and the converter handles NVIDIA's official BF16 checkpoint despite its renamed tensors. MTP speculative decoding isn't wired in yet, but the model runs today on plain builds.