# Sources

## TeichAI — native Qwen3.8 Fable LoRA

https://huggingface.co/TeichAI/Qwen3.8-27B-Fable-Distill-LoRA

Verified facts:

- adapter name: `Qwen3.8-27B-Fable5-Distill-LoRA`
- architecture tag: `qwen3_5`
- license: Apache-2.0
- model tree base: `Qwen/Qwen3.8-27B`
- cited public datasets:
  - `armand0e/claude-fable-5-claude-code`
  - `armand0e/Fable-5-Chat`
- TeichAI also reports use of personal/private Fable 5 data.

## TeichAI — merged Qwen3.8 Fable model

https://huggingface.co/TeichAI/Qwen3.8-27B-Fable-Distill

Published benchmark comparison:

| Model | ARC Challenge | ARC Easy | BoolQ |
|---|---:|---:|---:|
| Qwen3.8-27B | 0.591 | 0.782 | 0.896 |
| Qwen3.8-27B-Fable5-Distill | 0.637 | 0.832 | 0.911 |

The model card calls it a light tune on Qwen3.8 and documents:

- `enable_thinking`
- `reasoning_effort = low | medium | xhigh`
- default template reasoning effort: `xhigh`
- sampling: temperature 1.0, top_p 0.95, top_k 20

## TeichAI — official GGUF conversion

https://huggingface.co/TeichAI/Qwen3.8-27B-Fable-Distill-GGUF

Useful reference details:

- Qwen3.8 Fable GGUFs
- IQ4_XS listed around 16 GB in TeichAI's own build
- Q8_0 around 29 GB
- BF16 around 55 GB
- separate multimodal projector
- F16/BF16/F32 mmproj variants
- MTP/nextn head retained as an extra BF16 layer in TeichAI's build
- `qwen35.block_count = 65` in their MTP-preserving GGUF
- architecture reported as qwen35

## Blackfrost foundation used by this custom release

https://huggingface.co/Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16

This custom build uses the Blackfrost BF16 abliterated Qwen3.8 checkpoint as the merge foundation.

## llama.cpp

https://github.com/ggml-org/llama.cpp

Used for GGUF conversion, imatrix generation, quantization and inference.

## Final imatrix

NOT YET DOCUMENTED.

Do not insert chunk count, token count, hashes or tensor statistics until the exact post-merge imatrix used by the published IQ4_XS has been inspected.