---

license: apache-2.0

tags: [gguf, benchmark, mxfp4]

---

PXA benchmark files

Competitor-side model files used in the PXA cross-engine comparisons (bench/fair-battle.md),

published so every row can be reproduced from a download. Each is a stock model converted with the repo's own llama-quantize.

| File | Format | Size | sha256 (first 16) | Used in |

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

| qwen36-35b-MXFP4.gguf | MXFP4 | 19.1 GB | 5fc152ba23f958ef | MoE row, 2x V100 (35B-A3B class), expert-codec delta |

Full sums in checksums.sha256. Protocol: bench/fair/protocol.md in the engine repo.

Measured artifacts behind the PXA README rows

The single-card rows on the PXA front page were measured on the

two files below and on nothing else. They carry the names they have in bench/fair/weights/MANIFEST.sha256,

so a download here is byte-for-byte the file a published number came from.

| File | sha256 | Which published row it backs |

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

| fusion2-35b-U16-q8head.gguf | 643e55b07faf8b9b359fcb050494ec9c56e51c2db8d81751b4120b97782c5779 | The 1× Tesla P100 and 1× Tesla V100 single-card rows: decode 62.4 t/s and prefill 827–843 t/s on the P100, decode 101.3 t/s and prefill ~1,800–1,900 t/s on the V100. Recipes in docs/COOKBOOK.md; the launcher's 1× P100 (sm_60) and 1× V100 (sm_70) PXQU-16 rows in docs/LAUNCHER.md. Also the PRODUCT arm of bench/fair/run.sh. |

| PXA-Fusion2-35B-PXQ2.gguf | c11b45ef4b6c2301137f07d9039287e0b85e0c6eb26fd3cb736d1d0c8f0fee4d | All four 1× GTX 1080 Ti cells against upstream ik_llama.cpp — cold prefill 1,333.5, chat prefill 734.1, cold decode 35.37, chat decode 64.23 t/s — in docs/data/gtx1080ti-vs-ik-2026-09-07.csv, and the 1080 Ti line of the per-card table in docs/COOKBOOK.md / docs/LAUNCHER.md. |

Both are PXA-codec files: they need the PXA engine, not stock llama.cpp.

The rest of the PXA-Fusion2-35B manifest

bench/fair/weights/MANIFEST.sha256 in the engine repo lists four more Fusion2 artifacts. None of

them backs a published single-card speed row — they are here because the manifest names them, and

because the recipes in docs/COOKBOOK.md and the SWE-bench Lite reproduction in bench/swe/ call

for them. I put them in this repo so the whole manifest resolves from one place instead of half of

it pointing at a repo name that now belongs to a different model.

| File | Size | sha256 | What it is |

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

| PXA-Fusion2-35B-PXQ3.gguf | 14,696,020,352 B | e3de17ac1e3b49fd0f7ed2cb252c13e996b66442fbc44d2a9c481b0426ec7fbb | The 3-bit tier. Manifest artifact, no published single-card row. |

| PXA-Fusion2-35B-PXQ6.gguf | 18,722,552,256 B | 4e63c3fd0349ef698ff919d3ef93ad40f6644c6acd87cd7efc969cfc064ddf98 | The 4-bit flagship. PXQ6 is the pre-rename display name — these bytes are the PXQ4 tier (docs/RENAME-MAP.md), and the published filename is deliberately not renamed. It is the flagship the SWE-bench Lite run in bench/swe/ serves. Manifest artifact, no published single-card speed row. |

| PXA-Fusion2-35B-PXQ6-MTP.gguf | 19,173,193,440 B | f08dc102b38c4a23ae546d9e007b8dad30e043e848480d29809d2262d0b17531 | The same 4-bit flagship with the MTP head, for --spec-type mtp. Manifest artifact, no published single-card row. |

| mmproj-fusion2-f16.gguf | 902,823,040 B | 556ff9a8a0ffa3ff924e8c1b8aefb6e16a3551abc8c9b38643233f69592df14d | The vision projector for the Fusion2 tiers (--mmproj). Manifest artifact, no published single-card row. |

The .gguf tiers above are PXA-codec files: they need the PXA

engine, not stock llama.cpp, and they must be fully GPU-resident (-ngl 99).

Check what you downloaded

sha256sum -c checksums.sha256 --ignore-missing

Run it from the directory holding the files — the names in checksums.sha256 are bare. The same

sums are in bench/fair/weights/MANIFEST.sha256 in the engine repo, and bench/fair/run.sh

refuses to start if a file present on disk does not match.

I keep these here because the repo they were first published in, poisonxa/PXA-Fusion2-35B-GGUF,

was renamed to poisonxa/PXA-Fusion4-35B-GGUF — a different model on its main branch, though the

original bytes are still reachable there at revision 2d7d9e7e5f445b1ea38a5686b8cf9d6bc9aa29df.

An old link now lands on the rename, so the copies in this repo are the ones to use: they carry the

manifest's own filenames and the same hashes, and they are the files the published numbers were

measured on.