---

license: apache-2.0

base_model: unsloth/Qwen3.8-27B-NVFP4

pipeline_tag: text-generation

library_name: gguf

description: "EXPERIMENT. 8 NVFP4 GGUF tiers of Qwen3.8-27B carrying the community 'SSMFIX' conv1d rescale (layers 52-62 x alpha). Unverified idea, published for experimentation. Based on redashes' work and the ideas he credits."

tags:

- gguf

- nvfp4

- qwen3.8

- blackwell

- mtp

- speculative-decoding

- vision

- multimodal

- llama.cpp

- ssmfix

- experiment

language:

- en

- multilingual

---

Qwen3.8-27B-NVFP4-MTP-SSMFIX-GGUF

> [!WARNING]

> THIS IS AN EXPERIMENT, NOT AN IMPROVEMENT.

> These files apply an unverified community idea ("SSMFIX") to Qwen3.8-27B.

> Good results are NOT guaranteed. Bad or worse results are NOT unlikely.

> Several people have tried to validate this idea so far, with mixed results — see

> What we know so far before downloading anything.

> If you want the plain, non-experimental family, use

> esatapedico/Qwen3.8-27B-NVFP4-MTP-GGUF instead.

A family of eight NVFP4 GGUF files of Qwen3.8-27B (native VLM, MTP speculative head baked in, 262,144-token native context), each carrying a small, precisely-defined weight modification known in the community as the "SSMFIX": eight short SSM conv1d kernels in late layers are multiplied by per-layer scale factors (alphas).

The idea is not ours. It was proposed and documented by other people; this repository only packages their idea into ready-to-run GGUF tiers so that more users can test it on more hardware. All credit is theirs — see Credits.

Credits (all of it)

This repository exists only because of the following people and their public work:

1. redashes — redashes/Qwen3.8-27B-BF16-SSMFIX. The direct basis of these files: redashes turned the original analysis into a concrete recipe ("v2", per-layer strict alphas), shipped a patched BF16 checkpoint, ran benchmark batteries against the official weights, and documented everything on his model card. The layer set and alpha values below are redashes' v2 values, unchanged.
2. LuffyTheFox — the original analysis behind the idea ("Sig-ScaleSync"): identified inflated conv1d standard deviations in specific late SSM layers of Qwen3.8-27B and derived per-layer rescale factors. See Qwen/Qwen3.8-27B discussion #76 and the discussion pinned on LuffyTheFox's Genesis model card.
3. FGDumitru — qwen-ssm-repair: an independent, statistically-derived implementation of the same class of repair (MAD Z-score + peer-group median scaling), which converges on the same tail SSM layers. Referenced from redashes' card as part of the cross-validation of the diagnosis.
4. Everyone participating in the linked discussions who pushed back, reproduced numbers, corrected block-numbering mistakes, or provided counter-evidence — including the critical voices. Their contributions are exactly what makes the "mixed results" picture below honest.

We claim no credit for the idea, the diagnosis, or the recipe. Our contribution is limited to mechanical packaging: applying the published alphas inside GGUF files, re-tiering, verifying the bytes, and benchmarking on our own hardware.

Why this repository exists

At the time of publishing, the SSMFIX exists as (a) discussion posts with formulas and alphas, (b) redashes' BF16 safetensors checkpoint — a perfectly valid format, ready to use with transformers-style runtimes as-is — and (c) several community GGUF conversions of it, such as Luis23333/Qwen3.8-27B-SSMFIX-UD-Q3_K_XL-GGUF and grimoni/Qwen3.8-27B-SSMFIX-UD-Q4_K_XL-GGUF. We are not the first to package this idea as GGUF.

This repository is simply our own set of flavors of the same idea, published the way we publish our other families: the eight-tier size/precision ladder of our Qwen3.8-27B-NVFP4-MTP family — native NVFP4 backbone, vision projector and MTP speculative head intact — with the SSMFIX rescale applied. Those conversions serve llama.cpp and other GGUF-based engines already; ours exists so users of those engines can pick the idea up in our ladder without converting anything themselves.

What we know so far (honest state of the evidence)

Read this section as the summary; check the sources above for details.

The diagnosis (inflated conv1d scales in 8 late SSM layers) has independent support. Three separate implementations — LuffyTheFox's factor derivation, FGDumitru's statistical method, and redashes' packaging — converge on the same set of layers {52, 53, 56, 57, 58, 60, 61, 62}.

The fix's usefulness is genuinely unsettled:

• redashes' benchmarks (BF16 checkpoint, thinking off, temp 0) show TruthfulQA-gen +6~8 pp and small GSM8K/IFEval gains, at the cost of CMMLU −1.8 pp and, after correcting a broken harness, MT-Bench −0.19 vs official (the earlier "+0.42" claim was retracted when the evaluation bug was found).
• froggeric's counter-benchmark on the sibling Qwen3.6 models (284 data points across perplexity / needle-in-a-haystack / repetition, 4K–128K context) found perplexity degraded at every length with no NIAH or repetition improvement from the same class of conv1d repair.
• Anecdotal reports on redashes' release are positive but informal ("definitely improving"), and the original HF discussion contains a contested back-and-forth about whether the diagnosis is sound at all.
• No one has yet published long-context results. The idea's motivating failure mode (repetition loops / truncation at very long context) lives beyond ~75k tokens; none of the published evaluations covers that regime. Your guess is as good as ours.

Bottom line: treat these files as a hypothesis to test, not a fix to trust.

What exactly was changed

Eight tensors per file: blk.{52,53,56,57,58,60,61,62}.ssm_conv1d.weight, stored as F32, multiplied by the following factors (redashes' v2 values):

| Layer | Alpha | | Layer | Alpha |

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

| 52 | 0.5901 | | 60 | 0.4814 |

| 53 | 0.5548 | | 61 | 0.6533 |

| 56 | 0.5449 | | 62 | 0.6186 |

| 57 | 0.5357 | | | |

| 58 | 0.6097 | | | |

Every other tensor is untouched relative to the corresponding tier of the base family: the 448-tensor NVFP4 backbone is byte-identical to esatapedico/Qwen3.8-27B-NVFP4-MTP-GGUF, and mmproj-BF16.gguf is byte-identical to Unsloth's projector (it contains no SSM tensors).

How it was done: the alphas were applied directly inside the F32 ssm_conv1d tensors of the ORIG parent GGUF (lossless — these tensors are full-precision F32 in every file of this family), then all eight tiers were re-built from that parent with the same llama-quantize --tensor-type-file maps as the base family. Each modified tensor was verified bit-exact against (original × alpha) after quantization, and each file's embedded general.name/general.description metadata carries the SSMFIX marker, so a downloaded file self-identifies.

Parity caveat: because we did not start from redashes' patched safetensors checkpoint, we have not validated parity with such a version. The layer set and alpha values match his published recipe, and our own application is bit-exact by construction — but whether a GGUF converted from redashes' checkpoint would carry byte-identical weights to ours has not been checked. Treat the two lineages as expected-equivalent, not proven-equivalent.

The eight files

| File | Size | lm_head (output.weight) | token_embd | MTP head (blk.64) | Attention |

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

| Qwen3.8-27B-NVFP4-MTP-SSMFIX-VERY-LOW.gguf | 14.86 GB | Q3_K | Q2_K | Q2_K | NVFP4 |

| Qwen3.8-27B-NVFP4-MTP-SSMFIX-COMPACT-LOW.gguf | 15.16 GB | Q4_K | Q3_K | Q2_K | NVFP4 |

| Qwen3.8-27B-NVFP4-MTP-SSMFIX-LOW.gguf | 15.53 GB | Q5_0 | IQ4_XS | IQ4_XS | NVFP4 |

| Qwen3.8-27B-NVFP4-MTP-SSMFIX-MEDIUM.gguf | 16.38 GB | Q8_0 | Q6_K | IQ4_XS | NVFP4 |

| Qwen3.8-27B-NVFP4-MTP-SSMFIX-MID-HIGH.gguf | 16.91 GB | Q8_0 | Q8_0 | Q8_0 | NVFP4 |

| Qwen3.8-27B-NVFP4-MTP-SSMFIX-HIGH.gguf | 17.57 GB | BF16 | Q6_K | IQ4_XS | NVFP4 |

| Qwen3.8-27B-NVFP4-MTP-SSMFIX-VERY-HIGH.gguf | 19.69 GB | BF16 | BF16 | BF16 | NVFP4 |

| Qwen3.8-27B-NVFP4-MTP-SSMFIX-HIGHEST.gguf | 23.19 GB | Q8_0 | BF16 | BF16 | Q8_0 |

| mmproj-BF16.gguf | 931 MB | — (vision projector, unchanged) | | | |

SHA-256

ff3bca4026529cbc6b128a0b17ce84e9bbf0db796d60fad7cc267c7d89a46e08  Qwen3.8-27B-NVFP4-MTP-SSMFIX-VERY-LOW.gguf
c1b655f065ade8923164354b6e0fe07aef5ad18181c675aadd4df81e784d032b  Qwen3.8-27B-NVFP4-MTP-SSMFIX-COMPACT-LOW.gguf
42faa27add72d16f049c55413196aaf2f38341dd3a2e45d386234ddb247bd30a  Qwen3.8-27B-NVFP4-MTP-SSMFIX-LOW.gguf
54380e601e7301df215eaa37e97f807ec60c9bfd6c19aeb468a67cf0dc961b42  Qwen3.8-27B-NVFP4-MTP-SSMFIX-MEDIUM.gguf
a804ba9ea4bdc78c431f64547c6424d5ab831d02796f9b2be6b3a94c14d216c5  Qwen3.8-27B-NVFP4-MTP-SSMFIX-MID-HIGH.gguf
d445098f906fb4400a319d25adfd31b6bdbeb732318b7f2eb9af1e8b7ebb10c8  Qwen3.8-27B-NVFP4-MTP-SSMFIX-HIGH.gguf
fe1a3b0a72652c895fd46b6da429519023efead8de1ef882e73f65b08c85a155  Qwen3.8-27B-NVFP4-MTP-SSMFIX-VERY-HIGH.gguf
0215c011893680d2ae224eeb210644cdd81fb640cf47a329340290eca4eae862  Qwen3.8-27B-NVFP4-MTP-SSMFIX-HIGHEST.gguf
83ee4f4f205fa514161778c41df1ea14144faa0f713510893b63c2395f5c2d53  mmproj-BF16.gguf (byte-identical to unsloth's)

Usage

Identical to the base family — same flags, same requirements (recent llama.cpp with NVFP4 + Blackwell sm_120, --spec-type draft-mtp, --mmproj for vision):

llama-server \
  --model Qwen3.8-27B-NVFP4-MTP-SSMFIX-MEDIUM.gguf \
  --mmproj mmproj-BF16.gguf \
  --ctx-size 262144 \
  --flash-attn on \
  --spec-type draft-mtp \
  --spec-draft-n-max 6 \
  --spec-draft-p-min 0.75 \
  --temp 0.7 --top-p 0.95 --top-k 20

Benchmarks

Read this first: these are our own informal single-run numbers on one repetitive 180k-token payload — exactly the same naive setup as the base family's "first observations". They are not proof of anything. The model is non-deterministic; meaningful claims would need many repeated runs, which we have not done. Treat this as "what we saw once", nothing more.

Setup: 2x NVIDIA Blackwell 16 GB (RTX 5070 Ti + RTX 5060 Ti), split-mode: tensor, flash attention on, quantized KV cache; one fresh instance per tier; 180,052-token payload, max_tokens 20000, temp 0.6 / top_p 0.95 / top_k 20, MTP on (spec_n_max 6, spec_p_min 0.75). Baseline column = the unpatched tiers' published runs on the identical payload and harness (from the base family card).

| Tier | Prefill t/s (base → SSMFIX) | Decode t/s | Tokens generated (base → SSMFIX) | MTP acc | Finish |

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

| LOW | 645.6 → 627.9 | 16.9 | 1,151 → 659 | 0.749 | stop |

| MEDIUM | 649.3 → 641.6 | 22.9 | 1,907 → 383 | 0.820 | stop |

| HIGH | 647.6 → 644.6 | 17.4 | 1,795 → 1,946 | 0.734 | stop |

| VERY-HIGH | 650.4 → 642.7 | 16.2 | 2,007 → 910 | 0.723 | stop |

Repetition analysis (distinct-5-gram ratios, higher = less repetitive; flags = harness "suspicious" heuristic):

| Tier | Reasoning segment | Content segment |

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

| LOW | 0.948, clean | 0.996, clean |

| MEDIUM | 0.994, clean | 0.991, clean |

| HIGH | 0.877, flagged | 0.985, flagged |

| VERY-HIGH | 0.869, flagged | 1.000, clean |

What we saw, honestly:

• Prefill is unchanged within noise (~628-645 t/s everywhere) — expected, the NVFP4 backbone is byte-identical to the base family.
• Three of four tiers answered the same prompt with far fewer tokens than the unpatched tiers did (LOW −43%, MEDIUM −80%, VERY-HIGH −55%; HIGH was the exception at +8%). Directionally consistent with what the SSMFIX idea claims (less runaway reasoning), but again: one run each.
• MEDIUM had the cleanest output of any tier we have measured on this payload — including the unpatched ones. In the base family, MEDIUM showed the worst reasoning-segment repetition.
• HIGH and VERY-HIGH picked up reasoning-segment repetition flags that the unpatched tiers did not have. That is a real counter-signal, and it is why the warning at the top of this card says bad/worse outcomes are not unlikely.
• Decode t/s is not comparable across tiers/runs here because token counts differ wildly.

Net: genuinely mixed, which matches the community's experience with this idea so far. If you benchmark these files, please share your numbers — ideally repeated runs — on the linked discussions.

Reproduction

The exact per-tensor quantization maps used to build every tier from the patched parent are included in this repo (overrides-very-low.txt ... overrides-highest.txt, 1,202 entries each), same as in the base-family repo.

License

Apache-2.0, inherited from every upstream artifact (Qwen base model, Unsloth NVFP4 quantization, unsloth's projector). The patch applied here changes 64 KB of weights per file and introduces no new license terms.

"Qwen" is a trademark of Alibaba. This repository is not affiliated with, sponsored by, or endorsed by Alibaba, Unsloth, redashes, LuffyTheFox, or FGDumitru.

Note on this card

This model card was written by an AI assistant at the request of the repository author, who did the engineering. As with any AI-generated text, there may be errors; please verify anything important (hashes, sizes, commands) against the files themselves before relying on them.