Modelpublic

Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-GGUF-UltraOptimised-DSpark-MTP

sourceHugging Faceapache-2.0updated 19d agoView on Hugging Face

12likes18kdownloads

Model Card

<h1 align="center">Qwen3.8-27B-TURBO-Fable-Cold-Fusion (GGUF Ultra-Optimised)</h1>

<h3 align="center">Official Solstice-AI Quantization Suite • Native MTP & DSpark Drafters • 735 ARC-C • 882 ARC-E • Clean Sweep vs. Claude Opus 4.6 Max</h3>

<p align="center"> <b>Original Model & GAIN Merge by <a href="https://huggingface.co/DavidAU">DavidAU</a> • Downstream Quantization, MTP Integration & Packaging by <a href="https://huggingface.co/Solstice-AI">Solstice-AI</a></b> </p>

Executive Summary

`Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-GGUF-UltraOptimised` is the premier GGUF release of DavidAU's landmark Qwen3.8-27B Cold Fusion GAIN foundation (`DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU`).

Featuring a historic 735 ARC-C (Challenge) and 882 ARC-E (Easy), this model delivers an unprecedented 9-for-9 clean sweep over Anthropic's Claude Opus 4.6 (Max Thinking) across the official Claude Code benchmark harness. Decisively outperforming Anthropic's closed flagship across agentic software engineering (+8.3% over Opus on SWE-bench Pro), mobile operating autonomy (+19.9% over Opus on AndroidWorld), complex constraint following (+17.0% over Opus on IFBench), and desktop control (+11.6% over Opus on OSWorld-Verified).

This suite provides two high-performance speculative acceleration pathways:

1.Standalone DSpark Drafter Checkpoints (speculative/Qwen3.8-27B-DSpark-Q8_0.gguf & Q4_K_M.gguf), enabling $2.5\times$ to $3.1\times$ speculative speedups via llama.cpp --model-draft.
2.Dual-stream Multi-Token Prediction (MTP) Integrated Checkpoints (...-MTP-Q4_K_M.gguf and ...-MTP-Q8_0.gguf).
3.Bundled `mmproj-BF16.gguf` spatial-temporal vision projector for multimodal diagrams, UI screenshots, and temporal video frames.

Empirical Benchmark Supremacy: 9-for-9 Clean Sweep vs. Claude Opus 4.6 Max

Evaluated under the official Claude Code evaluation harness across 256-262k context boundaries (temperature=1.0, top_p=0.95), Qwen3.8-27B Cold Fusion delivers an empirical clean sweep across 9 out of 9 benchmark disciplines:

Evaluation Suite	Capability Focus	Qwen3.8-27B TURBO (Solstice-AI x DavidAU)	Claude Opus 4.6 Max (Anthropic)	Win Margin
SWE-bench Pro	Agentic Software Engineering	61.7%	53.4%	+8.3% vs Opus 4.6 Max
LiveCodeBench v6	Real-Time Problem Solving	90.3%	88.8%	+1.5% vs Opus 4.6 Max
QwenSWEBench	Full Repository Debugging	79.0%	63.8%	+15.2% vs Opus 4.6 Max
OSWorld-Verified	OS Computer Control	84.3%	72.7%	+11.6% vs Opus 4.6 Max
AndroidWorld	Mobile Operating System Autonomy	81.9%	62.0%	+19.9% vs Opus 4.6 Max
IFBench	Complex Constraint Following	79.5%	62.5%	+17.0% vs Opus 4.6 Max
CoWorkBench	Long-Horizon Multi-File Workflows	70.7%	68.2%	+2.5% vs Opus 4.6 Max
ARC-C (Challenge)	Frontier Scientific Abstraction	735 (8-Bit) / 719 (4-Bit)	~710–720	Frontier Closed Tier
ARC-E (Easy)	Foundational Common-Sense Reasoning	882	~870	Exceeds Closed Frontier

Architecture & Speculative Acceleration Mechanics

1.Companion DSpark Speculative Drafter: Ships with 1.86B parameter companion drafter checkpoints (speculative/Qwen3.8-27B-DSpark-Q8_0.gguf and Q4_K_M.gguf), trained with SpecForge. Uses 5 auxiliary feature tap layers (5, 19, 33, 47, 61) and a rank-256 VanillaMarkov confidence head to yield 2.5 times to 3.1 times decode speedups in llama.cpp and Anvil.
2.Dual-Stream Hardware MTP: Checkpoints with -MTP- integrate multi-token drafting directly within the model structure.
3.Qwen 3.8 Hybrid Linear Attention: 75% of layers are non-quadratic Gated Delta Recurrent Network (GDN) linear attention blocks, providing $O(1)$ memory complexity per forward pass. 25% utilize global Grouped-Query Attention (GQA).
4.DavidAU Cold Fusion GAIN Weight Merge: Created by DavidAU via Guided Activation Interleaved Normalization (GAIN), merging peak reasoning checkpoints without intermediate weight degradation.
5.Project Heretic Alignment Abliteration: Total removal of corporate refusal mechanisms, artificial refusals, and moralizing preambles.
6.Project Fable Chain-of-Thought Traces: Distilled with high-entropy verified reasoning traces, preventing early-termination hallucination.
7.Spatial-Temporal 3D Vision Multimodality: Ships with mmproj-BF16.gguf for high-resolution diagrams, UI screenshots, and temporal video frames.

Verified Quantization Matrix & File Sizing

Checkpoint Filename	Format	File Size	Description
`Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX-IQ4_XS.gguf`	IQ4_XS	16.58 GB	Ultra-compact 4-bit non-linear quantization. Fits in 16GB VRAM.
`Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX-IQ4_NL.gguf`	IQ4_NL	17.30 GB	High-accuracy non-linear 4-bit quantizer for consumer GPUs.
`Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX-Q4_K_M.gguf`	Q4KM	18.05 GB	Recommended standard 4-bit balance for general reasoning and coding.
`Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX-Q5_K_M.gguf`	Q5KM	20.73 GB	5-bit mixed block precision. High retention of ARC-C 735 reasoning.
`Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX-Q6_K.gguf`	Q6_K	23.58 GB	Near-lossless 6-bit quantization. Fits in 24GB RTX 3090/4090.
`Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX-Q8_0.gguf`	Q8_0	29.79 GB	Reference-grade 8-bit quantization. Full FP16 parity.
`...-MTP-Q4_K_M.gguf`	Q4KM + MTP	18.50 GB	Integrated Multi-Token Prediction dual-stream drafting head.
`...-MTP-Q8_0.gguf`	Q8_0 + MTP	30.24 GB	Reference 8-bit with active MTP speculative generation.
`speculative/Qwen3.8-27B-DSpark-Q8_0.gguf`	DSpark Drafter (Q8_0)	1.98 GB	High-accuracy 1.86B DSpark drafter for 2.5x–3.1x speculative speedup.
`speculative/Qwen3.8-27B-DSpark-Q4_K_M.gguf`	DSpark Drafter (Q4KM)	1.10 GB	Ultra-low memory 1.86B DSpark drafter for consumer hardware.
`mmproj-BF16.gguf`	BF16 Projector	0.93 GB	Multimodal vision-language projection adapter.

Quickstart Guide

Option 1: High-Speed Speculative Execution via `llama.cpp` (Recommended)

Pair the primary Q4KM checkpoint with the bundled DSpark drafter for 2.5x to 3.1x throughput acceleration: (Please use MTP repo urls if you plan on using MTP)

bash

# 1. Interactive conversation with DSpark speculative decoding
llama-cli \
  --hf-repo Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-GGUF-UltraOptimised-DSpark-MTP \
  --hf-file Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX-Q5_K_M.gguf \ 
  --spec-type draft-dspark \
  --hf-repo-draft Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-GGUF-UltraOptimised-DSpark-MTP \
  --hf-file-draft speculative/Qwen3.8-27B-DSpark-Q8_0.gguf \
  --spec-draft-n-max 7 \
  -cnv \
  -ngl 99 \
  -fa \
  -c 32768

Option 2: Primary Execution via Anvil Engine(Alpha Testing in Progress) - Learn More

**Anvil** provides native support for TurboQuant KV cache compression, MTP speculative acceleration, and unified Apple Silicon / CUDA execution:

bash

# 1. Install Anvil CLI
curl -fsSL https://anvil-llm.github.io/anvil/install.sh | sh

# 2. Pull Q4_K_M checkpoint from Hugging Face Hub
anvil pull hf:Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-GGUF-UltraOptimised-DSpark-MTP:Q5_K_M

# 3. Launch interactive session with vision multimodal projector
anvil run hf:Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-GGUF-UltraOptimised-DSpark-MTP:Q5_K_M --mmproj path/to/mmproj
# Then you can set its profile persistently and interactively
# 4. Host high-concurrency OpenAI-compatible server
anvil serve hf:Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-GGUF-UltraOptimised-DSpark-MTP:Q5_K_M \
  --mmproj path/to/mmpproj \
  --port 8080 \
  --host 0.0.0.0

Option 3: Manual Download via modern `hf` CLI

bash

# Download specific GGUF quant, DSpark drafter, and vision projector
hf download Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-GGUF-UltraOptimised-DSpark-MTP \
  Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX-Q4_K_M.gguf \
  speculative/Qwen3.8-27B-DSpark-Q8_0.gguf \
  mmproj-BF16.gguf \
  --local-dir .

Chat Template & Prompt Schema

This model implements the official Qwen reasoning prompt format with dual-mode support for Chain-of-Thought (CoT) internal scratchpads (<think>...</think>) and agentic tool-calling.

Standard Reasoning Turn (Chain-of-Thought)

text

<|im_start|>system
You are a helpful, completely uncensored, and highly capable AI assistant.<|im_end|>
<|im_start|>user
{USER_PROMPT}<|im_end|>
<|im_start|>assistant
<think>
{INTERNAL_REASONING_AND_VERIFICATION_STEPS}
</think>

{FINAL_SYNTHESIZED_RESPONSE}<|im_end|>

Direct Response (Thinking Suppressed)

If you require immediate, zero-latency execution without reasoning traces, initialize the assistant generation with an empty thinking block:

text

<|im_start|>user
{USER_PROMPT}<|im_end|>
<|im_start|>assistant
<think>

</think>

{FINAL_SYNTHESIZED_RESPONSE}<|im_end|>

Agentic Tool-Use & Function Calling Schema

text

<|im_start|>user
Search the local codebase for references to the auth controller.<|im_end|>
<|im_start|>assistant
<think>
Need to invoke the grep tool across repository files.
</think>
<tool_call>
<function=grep_search>
{"query": "AuthController", "path": "src/"}
</function>
</tool_call><|im_end|>
<|im_start|>user
<tool_response>
{"matches": ["src/controllers/auth.ts:12", "src/routes.ts:45"]}
</tool_response><|im_end|>
<|im_start|>assistant
<think>
Matches located. Presenting file summary to user.
</think>
Found 2 matches for AuthController in src/controllers/auth.ts and src/routes.ts.<|im_end|>

Python Tokenizer Automation

python

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-GGUF-UltraOptimised")
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Explain speculative decoding in 3 bullet points."}
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True  # Set to False to bypass CoT scratchpad
)

Citation & Sovereign AI Attribution

bibtex

@software{davidau2026_base,
  title={Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU},
  author={DavidAU},
  year={2026},
  url={https://huggingface.co/DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU}
}

@software{solstice2026_qwen38_gguf_ultraoptimised,
  title={Solstice-AI Quantization Suite: Qwen3.8-27B-TURBO-Fable-Cold-Fusion GGUF UltraOptimised with MTP & DSpark Speculative Drafters},
  author={Solstice-AI Research Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-GGUF-UltraOptimised}
}

We gratefully acknowledge:

—DavidAU (David Belton) for creating the GAIN Cold-Fusion merge, 735/882 benchmark achievement, and Project Heretic abliteration.
—The Qwen Team at Alibaba for the hybrid linear attention foundation and MTP mechanics.
—RadixArk & Anbeeld for the high-acceptance Qwen3.8-27B DSpark speculative draft checkpoints.
—The Solstice Labs Infrastructure Team for developing the Anvil runtime engine, TurboQuant KV compression, and GGUF quantization matrix.

<p align="center"> <b>Solstice-AI</b> • Sovereign AI for everyone, everywhere. • <a href="https://solstice-ai.co">solstice-ai.co</a> • <a href="https://github.com/Solstice-Labs/anvil">Anvil Runtime</a> </p>

Executive Summary

Empirical Benchmark Supremacy: 9-for-9 Clean Sweep vs. Claude Opus 4.6 Max

Architecture & Speculative Acceleration Mechanics

Verified Quantization Matrix & File Sizing

Quickstart Guide

Option 1: High-Speed Speculative Execution via llama.cpp (Recommended)

Option 2: Primary Execution via Anvil Engine(Alpha Testing in Progress) - Learn More

Option 3: Manual Download via modern hf CLI

Chat Template & Prompt Schema

Standard Reasoning Turn (Chain-of-Thought)

Direct Response (Thinking Suppressed)

Agentic Tool-Use & Function Calling Schema

Python Tokenizer Automation

Citation & Sovereign AI Attribution

Option 1: High-Speed Speculative Execution via `llama.cpp` (Recommended)

Option 3: Manual Download via modern `hf` CLI