---

license: other

license_name: nvidia-open-model-license

license_link: https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/

base_model: nvidia/Nemotron-3-Puzzle-75B-A9B

tags:

• gguf
• llama.cpp
• nemotron
• mamba2
• moe

---

Nemotron-3-Puzzle-75B-A9B — GGUF

First GGUF release of NVIDIA's Nemotron-3-Puzzle-75B-A9B (hybrid mamba2/attention/latent-MoE, 75B total / 9B active, 262k context, MTP draft head).

Converted from the official FP8 checkpoint (weight scales absorbed at conversion — no double quantization), then quantized from the Q8_0 master with an importance matrix.

Files

| file | size | note |

|---|---|---|

| Puzzle-75B-A9B-Q8_0-0000X-of-00002.gguf | 77.7 GiB (2 shards) | master, near-lossless — point llama.cpp at shard 00001, the rest loads automatically |

| Puzzle-75B-A9B-Q4_K_M-0000X-of-00002.gguf | 48.1 GiB (2 shards) | reference k-quant, fastest decode |

| Puzzle-75B-A9B-NVFP4.gguf | 45.0 GiB | experts NVFP4, everything else Q8_0 |

| Puzzle-75B-A9B-UD-IQ4-XL.gguf | 41.6 GiB | experts IQ4_XS; attn Q8_0, ssm/shexp Q6_K, ffn_latent Q8_0 |

| puzzle-imatrix.gguf | 0.2 GiB | reusable imatrix (calibration_datav3) |

Requirements

Not yet supported by mainline llama.cpp — needs per-layer heterogeneous MoE arrays and the 2-sub-block MTP head. Use the puzzle-port branch until the PR is merged: [PR_LINK]

Measured (Strix Halo 128GB unified, Radeon 8060S, -ngl 99; PPL = wikitext-2 test, 24 chunks)

| quant | PPL | decode t/s | prefill t/s | backend |

|---|---|---|---|---|

| Q8_0 | 5.325 | 10.2 | 189 | Vulkan |

| Q4_K_M | 5.404 | 19.9 | 238 | ROCm |

| UD-IQ4-XL | 5.377 | 17.7 | 211 | ROCm |

| NVFP4 | 5.383 | 16.6 | 243 | ROCm |

All three 4-bit variants sit within noise of each other on PPL (±0.08); pick by speed/size trade-off.

⚠️ On Strix Halo (gfx1151) use the ROCm/HIP backend for the 4-bit quants: Vulkan decode collapses to ~2.7 t/s on this model's MoE (mul_mat_id slow path). Q8_0 exceeds the ROCm allocation limit → run it on Vulkan.

MTP speculative decoding (--spec-type draft-mtp) loads and drafts correctly, but is currently slower than plain decoding (~13 vs 16.6 t/s): llama.cpp cannot yet roll back mamba2 recurrent states, which throttles draft attempts. Leave it off for now.

Notes

• Reasoning model: llama-server parses the thinking channel natively.
• AI-assisted work; everything reviewed and validated end-to-end on my hardware.