license: apache-2.0
language:
- en
- zh
pipeline_tag: image-text-to-text
library_name: transformers
tags: - solstice-ai
- anvil
- mlx
- mlx-my-repo
- 6-bit
- qwen3_8
- qwen3_6
- heretic
- uncensored
- cold-fusion
- gain-training
- multi-stage-tuning
- reasoning
- chain-of-thought
- token-efficient
base_model: DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU
model_name: Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored (MLX 6-Bit)
extra_gated_prompt: "This model is completely uncensored and de-aligned. You are solely responsible for outputs generated."
🌌 Qwen3.8-27B-TURBO-Fable-Cold-Fusion (6-Bit MLX)
Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored is an advanced frontier checkpoint combining GAIN Multi-Stage Post-Training, Cold-Fusion Weight Synthesis, Low-KLD Heretic De-alignment, and Precision Healing Datasets.
This repository provides the official 6-Bit Apple Silicon MLX quantization, delivering 99% of 8-bit cognitive performance with an ultra-compact memory footprint of ~22.4 GB VRAM, enabling full-speed local execution on Apple M-Series Silicon (M2/M3/M4 Pro, Max, and Ultra).
📊 Empirical Benchmarks vs. Base Qwen
The Stage 1b (735-882) post-heretic healing pipeline elevates core reasoning benchmarks 141 points above un-tuned Qwen 3.8-27B-Instruct while reducing reasoning token bloat by 2x to 10x:
| Model Checkpoint | ARC-C (0-shot) | ARC-E (0-shot) | BoolQ | HellaSwag | OpenBookQA | PIQA | WinoGrande |
|---|---|---|---|---|---|---|---|
| Qwen3.8-27B-TURBO-Fable (This Model, 6-Bit) | 0.735 | 0.882 | 0.917 | 0.832 | 0.530 | 0.837 | 0.785 |
| Qwen3.8-27B-Instruct (Baseline Base) | 0.591 | 0.782 | 0.896 | 0.746 | 0.448 | 0.801 | 0.711 |
| Qwen3.6-27B-Instruct | 0.647 | 0.803 | 0.910 | 0.773 | 0.450 | 0.806 | 0.742 |
| Qwen3.6-35B-A3B-Instruct (MoE) | 0.581 | 0.757 | 0.892 | 0.751 | 0.428 | 0.803 | 0.688 |
| Qwen3.5-27B-Instruct | 0.557 | 0.711 | 0.868 | 0.533 | 0.452 | 0.706 | 0.695 |
Key Takeaways:
- +14.4% ARC-Challenge jump over stock Qwen 3.8-27B-Instruct.
- +8.6% HellaSwag reasoning gain and +7.4% WinoGrande commonsense gain.
- Zero Alignment Degradation: Heretic decensoring executed with ultra-low KL-Divergence ($\text{KLD} = 0.0025$), followed by precision healing datasets.
⚡ Architecture & Training Highlights
- Multi-Stage GAIN Training: Multi-phase post-training aligning synthetic frontier reasoning traces without catastrophic forgetting.
- Dynamic Thinking & Token Efficiency: Automatically modulates
<thinking>chain-of-thought length based on prompt complexity (allocating deep scratchpads for hard math/code and direct responses for concise queries). - True Uncensored Capability: Complete removal of refusal vectors and corporate safety boilerplate, enabling unrestricted security research, red-teaming, and creative writing.
- Asymmetric MLX 6-Bit Quantization: Quantized with fused scale factors preserving critical attention projection layers in high precision.
🚀 Quickstart Guide
1. Run with mlx-lm (Python)
pip install mlx-lm
from mlx_lm import load, generate
model, tokenizer = load("Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-6Bit")
prompt = "Analyze the asymptotic memory complexity of Fast Walsh-Hadamard KV cache compression."
if hasattr(tokenizer, "apply_chat_template") and tokenizer.chat_template is not None:
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = generate(
model,
tokenizer,
prompt=prompt,
max_tokens=2048,
temp=0.7,
verbose=True
)
2. Local CLI Interactive Chat
mlx_lm.chat --model Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-6Bit
3. Launch OpenAI-Compatible Local Server
mlx_lm.server --model Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-6Bit --port 8080
💻 Hardware & Memory Requirements
| Apple Silicon Chip | Unified Memory | Generation Speed | Status |
|---|---|---|---|
| M2 / M3 / M4 Max (36GB - 64GB) | $\ge 32\text{ GB}$ | $38\text{--}46\text{ tok/s}$ | 🟢 Optimal |
| M1 / M2 / M3 / M4 Ultra (64GB - 192GB) | $\ge 64\text{ GB}$ | $62\text{--}78\text{ tok/s}$ | 🟢 Maximum Throughput |
| M3 / M4 Pro (24GB - 36GB) | $\ge 24\text{ GB}$ | $26\text{--}32\text{ tok/s}$ | 🟡 Supported |
📜 Attribution & Lineage
- Base Model Architecture:
Qwen/Qwen3.8-27B - Fine-Tuning & Multi-Stage GAIN: DavidAU & Nightmedia
- MLX Quantization & Optimization: Solstice-AI
- License: Apache 2.0
@misc{solstice2026qwen38turbo,
title={Qwen3.8-27B-TURBO-Fable-Cold-Fusion: Multi-Stage Distilled Uncensored Reasoning at 6-Bit MLX},
author={Solstice-AI Research and DavidAU and Nightmedia},
year={2026},
publisher={Hugging Face},
howpublished={\url{https://huggingface.co/Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-6Bit}}
}