Qwen3.8-27B Abliterated × Fable5-Distill — IQ4_XS - MTP ON
Abliterated Qwen3.8-27B + native Qwen3.8 Fable5-Distill behavior + post-merge imatrix IQ4_XS: a compact local build aimed at planning, coding, reasoning and agentic execution.
This release combines:
Foundation weights:Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16
LoRA:TeichAI/Qwen3.8-27B-Fable-Distill-LoRA
LoRA intended base:Qwen/Qwen3.8-27B
Architecture:qwen3_5, 27B class
Final format: GGUF
Target quant:IQ4_XS
Calibration: post-merge imatrix
This is not a cross-version Qwen3.6 → Qwen3.8 transplant. The TeichAI adapter is a native Qwen3.8-27B LoRA.
The only deliberate base variation is that the adapter is merged onto the Blackfrost abliterated derivative of Qwen3.8, rather than the untouched official Qwen3.8 checkpoint.
Why this build exists
The goal is straightforward: take the behavioral gains of TeichAI's Fable-distilled Qwen3.8 tune, preserve the lower-refusal character of the Blackfrost Qwen3.8 derivative, and package the result in a compact local format.
The result is aimed at users who want a 27B local model with strong planning, coding, long-form reasoning and agentic behavior, but without the memory footprint of BF16 or Q8.
Fable5-Distill: measured gains on Qwen3.8
TeichAI published direct benchmark results for its Qwen3.8-27B-Fable5-Distill model:
Benchmark
Qwen3.8-27B
Qwen3.8-27B Fable5-Distill
Gain
ARC Challenge
0.591
0.637
+0.046
ARC Challenge Easy
0.782
0.832
+0.050
BoolQ
0.896
0.911
+0.015
These are unusually clean results for a behavioral distill: the Fable tune improves all three published evaluation scores rather than trading one off against another.
TeichAI describes this as a light tune on Qwen3.8, trained using the cited public Fable 5 chat/agent datasets plus a substantially larger private corpus of personal Fable 5 data.
The adapter's Hugging Face model tree explicitly lists:
Based on TeichAI's model card and its direct Qwen3.8 benchmarks, the expected behavioral shift is toward:
stronger task decomposition;
more deliberate planning;
more agent-like execution;
better instruction retention across multi-step work;
stronger general reasoning signal on ARC;
improved straightforward factual/semantic decision accuracy on BoolQ;
preservation of Qwen3.8's thinking controls.
TeichAI's Qwen3.8 release explicitly supports:
text
1enable_thinking
2reasoning_effort = low | medium | xhigh
with the template defaulting to xhigh.
That makes this build attractive for coding agents, technical problem-solving, autonomous task planning, research workflows, and long-form implementation work.
Why the Blackfrost base
This model does not merge the Fable LoRA into vanilla Qwen3.8.
Instead, it starts from:
Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16
and merges the native Qwen3.8 Fable LoRA into those BF16 weights.
The resulting merged checkpoint is a new derivative and should be evaluated on its own. A benchmark from either parent is evidence about lineage, not automatically a measured score for the final IQ4_XS.
Expected strengths
Agentic coding
Qwen3.8 is designed around modern agentic workloads, and the Fable distill is trained on Fable 5 chat and agent traces.
The combination is particularly well suited to:
planning code changes before editing;
multi-file repository work;
debugging;
iterative implementation;
tool-oriented workflows;
implementation plans with dependencies and checkpoints.
Planning
This is the headline use case.
Fable-style data rewards a model for maintaining goals over multiple steps instead of immediately producing the first plausible answer.
Expect the model to be most interesting on prompts such as:
“inspect → plan → implement → verify”;
architecture decisions;
refactoring;
research plans;
multi-step technical tasks;
agent loops.
Reasoning
The TeichAI Qwen3.8 Fable model improved ARC Challenge from 0.591 to 0.637 and ARC Easy from 0.782 to 0.832.
That is concrete evidence that the tune does more than merely imitate a writing style.
Instruction comprehension
BoolQ improves from 0.896 to 0.911 in TeichAI's published comparison.
That supports the expectation that the Fable tune retains — and in this evaluation improves — basic instruction/question understanding.
Low-friction local use
The Blackfrost foundation was selected specifically for its abliterated / reduced-refusal character.
This can make the final model more useful for legitimate technical and research workflows that trigger unnecessary false-positive refusals in more conservative checkpoints.
IQ4_XS + post-merge imatrix
The final build uses IQ4_XS with a custom importance matrix generated after the Blackfrost Abliterated + TeichAI Fable5-Distill merge.
This is the exact imatrix artifact included in this repository:
llama_model_quantize_impl: have importance matrix data with 496 entries
Calibration volume recorded in the file:
280 × 512 = 143,360 token positions
This is calibration metadata, not a benchmark score. The imatrix guides llama.cpp's importance-aware quantization so that more sensitive activation directions receive preferential treatment during IQ4_XS compression.
The raw calibration corpus is not included here. The imatrix records its source as:
llama.cpp may intentionally keep selected sensitive tensors at higher precision. In the actual quantization log for this build, output.weight was converted to Q6_K while the overall model remained an IQ4_XS-class quant. This mixed-precision behavior is expected.
Multimodal / vision
TeichAI's Qwen3.8 Fable5-Distill is multimodal.
Their published GGUF build uses a separate mmproj and notes:
F16 projector: usual/default choice;
BF16 projector: matches source dtype;
F32 projector: fully unquantized projector.
This repository's primary IQ4_XS GGUF is the language model component.
For vision, publish and use a matching Qwen3.8 mmproj that you have actually tested.
