Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.
Nemotron Puzzle Fits 128GB on Paper. Loader Fix Landed. Not a GX10 Sit.
Official BF16 is 156,613,878,652 bytes; FP8 is 83,064,072,032; NVFP4 is 53,537,521,983. FP8 and NVFP4 fit 128GB on disk. The official card calls it 75.3B total and 9.3B active.
Key takeaways
- Official BF16 is 156,613,878,652 bytes; FP8 is 83,064,072,032; NVFP4 is 53,537,521,983. FP8 and NVFP4 fit 128GB on disk. The official card calls it 75.3B total and 9.3B active.
- Architecture is nemotron_h_puzzle: 88 blocks (40 Mamba2, 8 attention, 40 MoE). MoE is heterogeneous by layer. supports_mtp_export=False — skip MTP/NEXTN quants.
- llama.cpp PR 25444 merged the architecture. The expert-layer loader failure needed PR 28323 (n_expert_used_max), merged 2026-09-04T04:36:51Z. First tagged nightly with that fix is b10796 (published 2026-09-04T05:31:09Z). Latest checked: b10798. RemySkye Q4_K_M 51,601,807,488 / Q8_0 83,453,368,448 unchanged. I have not loaded this on the GX10. Not a sit.
Local LLMs on NVIDIA Spark / ASUS GX10
The headline is deliberately split in three. The puzzle fits 128GB on paper. The expert-layer loader fix landed in llama.cpp. I still have not sat it on the GX10.
The official sizes are the first receipt: BF16 is 156,613,878,652 bytes; FP8 is 83,064,072,032; NVFP4 is 53,537,521,983. FP8 and NVFP4 fit 128GB as an on-disk claim. BF16 does not. None of those numbers is leftover memory, context headroom, or a speed result.
The official card says 75.3B total / 9.3B active. License field: other. The license is the OpenMDW License Agreement 1.1. I am not laundering that into Apache or MIT.
The model type is nemotron_h_puzzle. There are 88 blocks: 40 Mamba2, 8 attention, and 40 MoE. The MoE layers are heterogeneous per layer, which matters here. supports_mtp_export=False. There are no MTP/NEXTN quants. Skip those. This is not a plain dense checkpoint with a familiar loader path.
The llama.cpp receipt
llama.cpp PR 25444 merged at 2026-09-03T06:53:09Z, commit c61b98b875eaa5e654a3f5c73b34c310d2c6ab4c. That got the architecture into the tree. It was not the same thing as a successful load. The loader errors that followed were:
model has expert layers but no expert layers are used
GGML_ASSERT(n_ids_used > 0) failed
PR 28323 (n_expert_used_max) closed that gap. It merged at 2026-09-04T04:36:51Z, merge commit 9a4843cf2f1a3fc8e39f8148e92ee6bfe18e2db6. The first tagged nightly carrying that fix is b10796, published 2026-09-04T05:31:09Z (target commit matches the merge). Latest nightly I checked is b10798, published 2026-09-04T15:03:27Z.
Two nearby receipts are real but do not replace this one. b10778 / PR 27483 reduces load peak RAM through per-tensor buffers. b10782 / PR 28198 adds concurrent streams per split for multi-GPU CUDA. Neither was the expert-count fix.
A merged loader fix is still not my GX10 sit. I have not loaded Nemotron Puzzle on this box, and I am not inventing leftover unified memory or tokens per second.
Community GGUFs are a separate line
RemySkye's listed community quants are unchanged in this check: Q4_K_M 51,601,807,488, Q8_0 83,453,368,448. Those are community conversions, not official NVIDIA artifacts. The sizes support an on-disk fit statement for those individual files. They do not supply a GX10 result.
Danbev manually converted and tested Q8_0 and Q4_K_M on a DGX Spark. There is no tok/s in this note and no logits verification. That is useful conversion evidence, not my GX10 bench. I am not quoting it as one.
So the status is narrow: official FP8, official NVFP4, and the listed community quants fit 128GB on disk. The expert-layer loader fix is in llama.cpp as of b10796. I have not sat the puzzle on the GX10. This is not a GX10 sit.
I run Grok when it earns it, and a GX10 when I want the weights in the room. This one is a size claim plus a loader ticket that closed. When I have actual GX10 receipts, that will be a different article.
Frequently asked questions
The official FP8 file is 83,064,072,032 bytes and the official NVFP4 file is 53,537,521,983 bytes, so both fit as an on-disk claim. The official BF16 file is 156,613,878,652 bytes and does not. Fit is not a run.
The expert-layer loader fix is in the tree. PR 28323 merged at 2026-09-04T04:36:51Z and first showed up in tagged nightly b10796. I have not run a load myself on the GX10, so I am not claiming a successful sit from this card.
No. danbev manually converted and tested Q8_0 and Q4_K_M on a DGX Spark, without tok/s or logits verification. That is not my GX10 bench, and I am not turning it into one.
Affiliate Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.