Downloads · 30 days
67
5% of all-time downloads
Verkiki/Qwen3.8-27B-F711-ColdGAIN-Fable5Fusion-MTP-GGUF
Qwen3.8-27B-F711-ColdGAIN-Fable5Fusion-MTP-GGUF is a image-text-to-text model from Verkiki. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. The card lists the license as other.
Not a generic Qwen quant. This release starts from Nightmedia's Fable-Fusion F711 GAIN MXFP4 checkpoint — a high-performing experimental fusion whose published MXFP4 source measurements reached 0.690 ARC, 3.803 ± 0.02…
Downloads · 30 days
67
5% of all-time downloads
All-time downloads
1.3K
Public
Repo size
16.3 GB
Likes
3
Public
Click a slice to open those files.
.gguf16.3 GB · 100%
From the Hugging Face model README
Not a generic Qwen quant. This release starts from Nightmedia's Fable-Fusion F711 GAIN MXFP4 checkpoint — a high-performing experimental fusion whose published MXFP4 source measurements reached 0.690 ARC, 3.803 ± 0.023 perplexity, and 185 tok/s. The exact MXFP4 checkpoint was then structurally recovered, dequantized into an ordinary FP16 bridge representation, calibrated with a model-specific importance matrix, converted to IQ4_XS, and completed with an embedded Qwen3.8-compatible MTP/NextN head.
The result is a compact GGUF release built around a model lineage that already showed a large measurable jump over stock Qwen3.8 in Nightmedia's own same-family MXFP4 tests — while preserving Qwen3.8's 27B / 64-layer multimodal architecture and local-runtime flexibility.
Link for script to recover mxfp4 to gguf: https://github.com/verkiki/quantbridge - this version was not tested yet; an older version was used to generate the f16.
Link for script to extract mtp from donor and put into the new gguf: https://github.com/verkiki/mtp-forge - this version was not tested yet; an older version was used to generate the final IQ4_XS.
Nightmedia's published MXFP4 measurements show a clear progression from stock Qwen3.8 through GAIN, Fable, and finally F711:
| Published MXFP4 model | ARC | Perplexity | Peak memory | tok/s |
|---|---|---|---|---|
| Qwen3.8-27B baseline | 0.581 | 5.952 ± 0.051 | 21.30 GB | 148 |
| Cold-Fusion GAIN V1.1 | 0.644 | 4.042 ± 0.026 | 21.30 GB | 185 |
| Cold-Fusion GAIN V1.1 + Fable | 0.654 | 3.975 ± 0.025 | 21.30 GB | 186 |
| Fable-Fusion F711 GAIN | 0.690 | 3.803 ± 0.023 | 21.30 GB | 185 |
Against the published stock Qwen3.8 MXFP4 baseline, the F711 source shows:
That is the performance signal this release is built on.
Benchmark honesty: those are measurements of Nightmedia's upstream MXFP4 F711 source, not newly measured scores for this IQ4_XS binary. This release does not relabel inherited results as fresh benchmarks.
The Fable component is not just a name in the merge tree. TeichAI reports the Qwen3.8 Fable Distill moving from the Qwen3.8 baseline to:
| Model | ARC Challenge | ARC Easy | BoolQ |
|---|---|---|---|
| Qwen3.8-27B | 0.591 | 0.782 | 0.896 |
| Qwen3.8-27B-Fable-Distill | 0.637 | 0.832 | 0.911 |
That corresponds to approximately +7.8% relative ARC, +6.4% relative ARC Easy, and +1.7% relative BoolQ in the published Fable evaluation.
Qwen3.8-27B is a dense 27B, 64-layer native vision-language architecture with a 262,144-token native context window, extensible by Qwen to 1,000,000 tokens, and a foundation trained with Multi-Token Prediction. It supports image/video understanding and flexible reasoning control at the architecture/template level.
This release keeps the multimodal model path. Vision is used through the matching mmproj GGUF.
This is important: the available Nightmedia source was already MXFP4.
This release was not produced from an unreleased original BF16 checkpoint. The FP16 checkpoint used later in the workflow was a reconstruction of the exact MXFP4 source values and served as a bridge format for GGUF tooling.
Nightmedia Fable-Fusion F711 GAIN
BF16 merge lineage
↓
Nightmedia MLX MXFP4 release
↓
structural shard validation
↓
safe tensor I/O + MLX MXFP4 dequantization
↓
ordinary FP16 reconstruction of the available MXFP4 values
↓
model-specific importance-matrix calibration
↓
IQ4_XS GGUF backbone
↓
embedded Qwen3.8-compatible MTP/NextN graft
↓
FableFusion GAIN F711 27B — final GGUF
The local source checkpoint contained 3 Safetensors shards:
All three shard headers and tensor byte ranges were validated before conversion.
The source tensor plan contained:
The source quantization metadata was:
Before the bulk conversion, a real attention projection was tested end-to-end:
packed weight: (1024, 640) uint32
scales: (1024, 160) uint8
↓ MXFP4 dequantization
FP16 weight: (1024, 5120)
This is why the release is described as an MXFP4-recovered IQ4_XS, not as a simple "BF16 → Q4" conversion.
The source files were not corrupt. Standard Safetensors access successfully validated and exposed all source entries, while MLX's direct large-file loading path failed on Windows. The recovery therefore used normal Safetensors I/O for the files and MLX specifically for the MXFP4 dequantization math.
Plain BF16 tensors that could not be exposed through NumPy were handled from their serialized payload and converted to FP16 explicitly.
The importance matrix was generated from a calibration representation of this recovered F711 backbone rather than copied from an unrelated public quant.
The final IQ4_XS therefore uses calibration data tied to the same recovered model being released.
The Nightmedia MXFP4 source declared MTP in configuration but contained no physical MTP tensor block in the audited weights.
The final GGUF therefore adds an embedded compatible Qwen3.8 NextN/MTP block after the F711 backbone quantization. The donor block was copied in GGUF form rather than being treated as part of the recovered Nightmedia source.
The result is a single main GGUF designed for current llama.cpp builds that support:
--spec-type draft-mtp
No separate draft-model file is required when the embedded MTP block is recognized by the runtime.
The final Nightmedia F711 source is a merge of merges, not a two-model fine-tune.
Nightmedia documents Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-Fable as a BF16 NuSLERP merge of:
DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1 weight 1.6
armand0e/Qwen3.8-27B-Fable-Distill weight 0.4
merge method: NuSLERP
dtype: bfloat16
The Fable model is currently published under the TeichAI account as TeichAI/Qwen3.8-27B-Fable-Distill; Nightmedia's recipe records the component under the armand0e name.
DavidAU describes the Cold Fusion model as a light, focused Qwen3.8 tune built with the GAIN + Unsloth training approach, targeting stronger general capability and shorter reasoning traces.
Nightmedia's Qwen3.6-27B-Architect-Polaris2-Fable-B-F451 is itself a BF16 NuSLERP merge:
Qwen3.6-27B-Architect-Polaris2-Fable-B weight 1.4
Qwen3.6-27B-Architect-Polaris-Fable-F451 weight 0.6
merge method: NuSLERP
dtype: bfloat16
Nightmedia reports for this branch, at MXFP4:
ARC 0.701
ARC Easy 0.873
BoolQ 0.909
HellaSwag 0.786
OpenBookQA 0.488
PIQA 0.813
WinoGrande 0.759
PPL 3.854 ± 0.024
tok/s 197
Nightmedia then combines the Qwen3.8 GAIN+Fable branch with the Qwen3.6 Architect/Polaris/Fable/F451 branch:
Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-Fable weight 1.4
Qwen3.6-27B-Architect-Polaris2-Fable-B-F451 weight 0.6
merge method: NuSLERP
dtype: bfloat16
That merge became the Fable-Fusion F711 GAIN lineage that Nightmedia later released in MLX MXFP4 form — the exact source recovered for this GGUF.
Qwen/Qwen3.8-27B
│
├── DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1
│
└── armand0e / TeichAI Qwen3.8-27B-Fable-Distill
│
└── NuSLERP BF16 (1.6 / 0.4)
↓
Nightmedia Qwen3.8 Cold-Fusion-GAIN-V1.1-Fable
│
│ weight 1.4
│
├───────────────────────────────────────┐
│
Qwen3.6 Architect / Polaris / Fable / F451 lineage │
│ │
├── Architect-Polaris2-Fable-B │
└── Architect-Polaris-Fable-F451 │
│ │
└── NuSLERP BF16 (1.4 / 0.6) │
↓ │
Nightmedia Qwen3.6 Architect-Polaris2-Fable-B-F451 │
│ │
│ weight 0.6 │
└───────────────────────────────────────┘
↓
NuSLERP BF16
↓
Nightmedia Fable-Fusion F711 GAIN
↓
MLX MXFP4 release
↓
Verkiki: recovery + custom imatrix + IQ4_XS
↓
embedded MTP / NextN block
↓
FableFusion GAIN F711 27B
This release is aimed at people who want one strong 27B local model that does more than generic chat:
mmproj;llama.cpp builds;The strongest quantitative evidence available before a dedicated benchmark of this exact IQ4_XS binary is the upstream F711 source benchmark ladder shown above. The final IQ4_XS should be evaluated separately if you need apples-to-apples numbers for your own hardware.
FableFusion-GAIN-F711-27B-IQ4_XS-MTP.gguf
mmproj-FableFusion-GAIN-F711-27B-F16.gguf
The first file is the main model. The second enables image/video input in GGUF runtimes that use an external multimodal projector.
llama-cli \
-m FableFusion-GAIN-F711-27B-IQ4_XS-MTP.gguf \
-c 8192 \
-p "Analyze the trade-offs between speculative decoding and standard autoregressive decoding."
llama-cli \
-m FableFusion-GAIN-F711-27B-IQ4_XS-MTP.gguf \
--spec-type draft-mtp \
--spec-draft-n-max 2 \
-c 8192 \
-p "Design a robust multi-stage research workflow and justify each stage."
Use the matching projector:
llama-mtmd-cli \
-m FableFusion-GAIN-F711-27B-IQ4_XS-MTP.gguf \
--mmproj mmproj-FableFusion-GAIN-F711-27B-F16.gguf \
--image image.jpg \
-p "Describe the image and explain the evidence behind your conclusions."
Use a recent llama.cpp build for Qwen3.8, multimodal and embedded-MTP support.
This model exists because of a long chain of work. Please preserve attribution when mirroring or deriving from it.
For redistribution and provenance notes, see NOTICE.md.
The benchmark values in this card are transcribed from the upstream model cards linked above.
They should be interpreted as lineage evidence:
That distinction matters — and it is why the numbers here are useful rather than misleading.