---
license: apache-2.0
base_model:
  - Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16
  - TeichAI/Qwen3.8-27B-Fable-Distill-LoRA
library_name: llama.cpp
pipeline_tag: text-generation
tags:
  - gguf
  - qwen3.8
  - qwen3_5
  - 27b
  - fable-5
  - fable-distill
  - abliterated
  - iq4_xs
  - imatrix
  - coding
  - agents
  - reasoning
  - multimodal
---

# 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:

```text
Base model: Qwen/Qwen3.8-27B
→ TeichAI/Qwen3.8-27B-Fable-Distill-LoRA
```

So this release keeps the LoRA on the same Qwen3.8 model generation for which it was trained.

**Source:**  
`TeichAI/Qwen3.8-27B-Fable-Distill-LoRA`  
`TeichAI/Qwen3.8-27B-Fable-Distill`

---

# What Fable5-Distill adds

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
enable_thinking
reasoning_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:

```text
Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16
```

and merges the native Qwen3.8 Fable LoRA into those BF16 weights.

The intended effect is to combine:

```text
Blackfrost Qwen3.8
    ↓
reduced refusal tendency

+

TeichAI Qwen3.8 Fable5-Distill LoRA
    ↓
measured reasoning / instruction benchmark gains
    ↓
Fable-style planning and execution behavior
```

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:

```text
filename:            Qwen3.8-27B-Abliterated-Fable5-Distill-imatrix.gguf
GGUF version:        3
general.type:        imatrix
GGUF tensor count:   992
imatrix entries:     496
chunk_count:         280
chunk_size:          512
token positions:     143,360
dataset recorded:    C:\AI\calibration.txt
file size:           13,642,624 bytes
SHA-256:             82ca0c264e505aae0dc391f31d8e447281085cee8ab4314c835ae4edc18b5db2
```

The quantizer log for this build confirmed:

```text
llama_model_quantize_impl: have importance matrix data with 496 entries
```

Calibration volume recorded in the file:

```text
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:

```text
C:\AI\calibration.txt
```

## Not a blind 4-bit conversion

```text
Blackfrost Qwen3.8 Abliterated BF16
        +
TeichAI Qwen3.8 Fable5-Distill LoRA
        ↓
merged high-precision Qwen3.8
        ↓
post-merge calibration
        ↓
496-entry importance matrix
        ↓
IQ4_XS
```

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.

Example:

```bash
llama-server \
  -m Qwen3.8-27B-Abliterated-Fable5-Distill-IQ4_XS.gguf \
  --mmproj mmproj-F16.gguf
```

---

# MTP / NextN

The official TeichAI Qwen3.8 Fable GGUF release retains the Multi-Token Prediction head:

```text
64 transformer layers
+ 1 MTP / nextn layer
= qwen35.block_count 65
```

TeichAI keeps that MTP head in BF16 in its own quantizations.

If this custom release is converted with MTP removed (`--no-mtp` / `--no-nextn`), do **not** claim MTP speculative decoding support.

The final repository should state the exact build choice after the final GGUF is verified.

---

# Sampling

TeichAI lists the Qwen3.8 base recommendations as:

```text
temperature = 1.0
top_p       = 0.95
top_k       = 20
```

Thinking controls:

```text
enable_thinking = true/false
reasoning_effort = low | medium | xhigh
```

For coding and agent workflows, `medium` is a practical place to start; use `xhigh` when the task benefits from longer deliberation.

---

# llama.cpp

```bash
llama-cli \
  -m Qwen3.8-27B-Abliterated-Fable5-Distill-IQ4_XS.gguf \
  -c 8192
```

Server:

```bash
llama-server \
  -m Qwen3.8-27B-Abliterated-Fable5-Distill-IQ4_XS.gguf \
  -c 8192
```

---

# LM Studio

Import:

```powershell
lms import "Qwen3.8-27B-Abliterated-Fable5-Distill-IQ4_XS.gguf"
```

Then configure GPU offload according to available VRAM.

---

# Benchmark integrity

This repository makes a strict distinction between **measured parent-model evidence** and **measured final-model evidence**.

### Directly measured by TeichAI on Qwen3.8 Fable5-Distill

- ARC Challenge: **0.637**
- ARC Easy: **0.832**
- BoolQ: **0.911**

### Baseline Qwen3.8 numbers from the same TeichAI comparison

- ARC Challenge: 0.591
- ARC Easy: 0.782
- BoolQ: 0.896

### Not yet claimed for this custom Blackfrost + Fable IQ4_XS

- those exact benchmark scores;
- final imatrix statistics;
- final refusal rate;
- final vision benchmark;
- final long-context benchmark.

Once the published IQ4_XS is benchmarked directly, its results should be added in a separate **Final Build Benchmarks** section.

---

# Credits

- **Qwen Team / Alibaba** — `Qwen/Qwen3.8-27B`
- **Blackfrost-AI** — `Blackfrost-AI/Qwen3.8-27B-ABLITERATED-BF16`
- **TeichAI** — `TeichAI/Qwen3.8-27B-Fable-Distill-LoRA`
- **armand0e** — `claude-fable-5-claude-code` and `Fable-5-Chat`
- **@nightmedia** — benchmark work credited by TeichAI
- **Unsloth + Hugging Face TRL** — training stack credited by TeichAI
- **ggml-org / llama.cpp** — GGUF, imatrix, IQ4_XS and runtime tooling

This is an independent derivative and is not an official release or endorsement by any of the above projects.

---

# Sources

See `SOURCES.md`.