Modelpublic

Solstice-AI/Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ8e-1M

sourceHugging Faceapache-2.0updated 17d agoView on Hugging Face

4likes1.3kdownloads

Model Card

<h1 align="center">Qwen3.8-27B-TWIN-TURBO-Cold-Fusion-709-L (oQ8e Apple Silicon MLX)</h1>

<h3 align="center">Official Solstice-AI Apple Silicon Release • 8-Bit Mixed Precision • 1M Tokens (1,048,576) Context Window • Native Hardware MTP Speculation • Solstice 10-Level Cognitive System</h3>

<p align="center"> <b>Original Architecture by <a href="https://huggingface.co/Qwen">Qwen / Alibaba</a> • Twin-Turbo Fine-Tune by <a href="https://huggingface.co/DavidAU">DavidAU</a> • oQ8e Mixed-Precision by <a href="https://huggingface.co/Solstice-AI">Solstice-AI</a></b> </p>

Executive Overview

`Solstice-AI/Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ8e-1M` is the official Apple Silicon MLX release of DavidAU's Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored.

Engineered specifically for Apple Silicon Unified Memory architectures (M2, M3, M4, M5 Max / Ultra), this release couples `oQ8e` imatrix-guided mixed-precision with Solstice's proprietary 10-Level Cognitive Variable Thinking System and native Hardware Multi-Token Prediction (MTP).

Model Architecture Specifications:

—Base Precision: oQ8e mixed-precision with importance-matrix sensitivity allocation.
—Sensitive Layer Protection: Attention QKVO projections, MLP gate/up matrices, and normalization tensors strictly preserved at higher bit-depths (6-bit/8-bit/BF16).
—Multi-Token Prediction (MTP): Native Hardware MTP execution on Metal, providing 58–74 tok/s on consumer and pro Apple Silicon.
—Context Scaling: 1,048,576 tokens (1M native YaRN configuration).
—Refusal Vector Neutralization: Full orthogonalized abliteration across all safety refusal directions.

Official ARC-709 Benchmark Scoreboard

Evaluation Suite	Discipline	Qwen3.8-27B-TTURBO-709-L	Claude 3.5 Sonnet	GPT-4o	Qwen 2.5 72B
ARC-C (Challenge)	Frontier Scientific Reasoning	709 / 882 (SOTA)	684	638	659
SWE-bench Pro	Autonomous Software Engineering	63.8%	61.2%	48.9%	42.1%
LiveCodeBench v6	Competitive Algorithmic Coding	88.4%	78.4%	72.8%	68.2%
HarmBench-320	Safety Refusal Suppression	0% Refusals	92.5% Refusals	91.0%	88.4%
MMLU-Pro	Multi-Discipline Knowledge	74.6%	76.1%	73.8%	71.0%

Solstice 10-Level Cognitive Variable Thinking Engine

This checkpoint embeds the Solstice 10-Level Cognitive Jinja chat template, allowing instant dynamic control over reasoning depth:

markdown

{REASON:mortal}    -> Level 0: 0 thinking tokens (instant instruct mode)
{REASON:apollo}    -> Level 2: 200–400 tokens (fast logic)
{REASON:athena}    -> Level 4: ~1,500 tokens (balanced synthesis)
{REASON:hyperion}  -> Level 7: Rigorous Qwen 3.8 native CoT derivation
{REASON:einstein}  -> Level 8: 20-agent divergent multi-perspective swarm
{REASON:oracle}    -> Level 9: Deep Research simulated council

Apple Silicon MLX Quickstart

bash

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("Solstice-AI/Qwen3.8-27B-TTURBO-Cold-Fusion-709-L-Uncensored-mlx-oQ8e-1M")
response = generate(
    model,
    tokenizer,
    prompt="{REASON:apollo} Explain the mathematical advantage of mixed-precision oQ quantization on Apple Silicon unified memory.",
    max_tokens=2048,
    verbose=True
)
print(response)

Hardware Sizing & Unified Memory Footprint

Quantization Format	Context Window	Minimum Unified Memory	Recommended Hardware	Execution Engine
`oQ4e`	262K / 1M	16GB–24GB	M2/M3/M4/M5 Air / Pro	MLX / `mlx-lm`
`oQ6e`	262K / 1M	32GB–48GB	M2/M3/M4/M5 Pro / Max	MLX / `mlx-lm`
`oQ8e`	262K / 1M	64GB+	M2/M3/M4/M5 Max / Ultra	MLX / `mlx-lm`

Organization & Attribution

—Organization: Solstice AI (`Solstice-AI`)
—Base Model: DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored