Solstice-AI

Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-6Bit

README

License: apache-2.0

📊 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:

Table with columns: Model Checkpoint, ARC-C (0-shot), ARC-E (0-shot), BoolQ, HellaSwag, OpenBookQA, PIQA, WinoGrande
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 ( $KLD = 0.0025$ ), 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)

bash
pip install mlx-lm

python
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

bash
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

bash
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

Table with columns: Apple Silicon Chip, Unified Memory, Generation Speed, Status
Apple Silicon Chip	Unified Memory	Generation Speed	Status
M2 / M3 / M4 Max (36GB - 64GB)	$\geq 32 GB$	$38 - 46 tok/s$	🟢 Optimal
M1 / M2 / M3 / M4 Ultra (64GB - 192GB)	$\geq 64 GB$

📜 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

bibtex
@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}}
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Solstice-AI

Model Tree

Base

DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU

Quantized

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

📊 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:

Table with columns: Model Checkpoint, ARC-C (0-shot), ARC-E (0-shot), BoolQ, HellaSwag, OpenBookQA, PIQA, WinoGrande
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 ( $KLD = 0.0025$ ), 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)

bash
pip install mlx-lm

python
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

bash
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

bash
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

Table with columns: Apple Silicon Chip, Unified Memory, Generation Speed, Status
Apple Silicon Chip	Unified Memory	Generation Speed	Status
M2 / M3 / M4 Max (36GB - 64GB)	$\geq 32 GB$	$38 - 46 tok/s$	🟢 Optimal
M1 / M2 / M3 / M4 Ultra (64GB - 192GB)	$\geq 64 GB$

📜 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

bibtex
@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}}
}

Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-6Bit

README

📊 Empirical Benchmarks vs. Base Qwen

⚡ Architecture & Training Highlights

🚀 Quickstart Guide

1. Run with mlx-lm (Python)

2. Local CLI Interactive Chat

3. Launch OpenAI-Compatible Local Server

💻 Hardware & Memory Requirements

📜 Attribution & Lineage

Explore FriendliAI today

README

📊 Empirical Benchmarks vs. Base Qwen

⚡ Architecture & Training Highlights

🚀 Quickstart Guide

1. Run with mlx-lm (Python)

2. Local CLI Interactive Chat

3. Launch OpenAI-Compatible Local Server

💻 Hardware & Memory Requirements

📜 Attribution & Lineage

1. Run with `mlx-lm` (Python)

1. Run with `mlx-lm` (Python)