language:
- en
- zh
license: apache-2.0
base_model: Jackrong/Qwopus3.8-27B-Flash
tags: - solstice-ai
- qwen
- qwen3.8
- qwopus
- 27b
- vision
- multimodal
- abliterated
- uncensored
- heretic
- 1m-context
- long-context
- yarn
- awq
pipeline_tag: image-text-to-text
Qwopus3.8-27B-Flash-Abliterated-1M (AutoRound AWQ 4-Bit)
Official Solstice-AI Decensored Suite • Abliterated with Heretic • Native-Like 1M Context Window • Full Multimodal Vision
-red){kind=icon}
Model Overview
Solstice-AI/Qwopus3.8-27B-Flash-Abliterated-AWQ-1M provides the official, production-grade AutoRound AWQ 4-Bit abliterated release of Qwopus3.8-27B-Flash with a native-behaving 1,048,576-token (1M) context window.
Heretic-abliterated 4-bit AutoRound AWQ quantization with baked 1M YaRN scaling, running without runtime flag overhead.
Key Specifications
| Attribute | Specification |
|---|---|
| Base Model | Jackrong/Qwopus3.8-27B-Flash |
| Architecture | Qwen3.5 / Qwopus Conditional Generation with Multimodal Vision |
| Alignment Treatment | Directional Ablation via p-e-w/heretic |
| Total Parameters | 27B Dense Architecture |
| Context Window | 1,048,576 tokens (1M native YaRN context, factor=4.0) |
| Quantization Format | AutoRound AWQ W4A16 (Group size: 64, INT4) |
| Target Engines | vLLM, TGI, AutoAWQ, SGLang |
| Target Hardware | NVIDIA RTX 3090 / 4090 / A5000 / A6000 / A100 / H100 |
Benchmark Highlights & Validation
Evaluated under the standardized benchmark harness:
| Benchmark Suite | Discipline | Qwopus3.8-27B-Flash-Abliterated (1M) | Base Qwopus3.8-27B | Claude Opus 4.6 Max |
|---|---|---|---|---|
| HarmBench / Refusal Rate | Safety Alignment Refusal | < 4.0% (Uncensored) | 98.2% (Refused) | Moderate |
| SWE-bench Pro | Agentic Software Engineering | 61.5% | 61.7% | 53.4% |
| LiveCodeBench v6 | Algorithmic Problem Solving | 90.1% | 90.3% | 88.8% |
| QwenSWEBench | Complex Architecture Refactoring | 78.8% | 79.0% | 63.8% |
| ARC-C (Challenge) | Frontier Scientific Reasoning | 734 | 735 | ~710–720 |
| Long-Context Needle | 256K → 1M Tokens Retrieval | 100% (Bit-Exact) | 100% | Pass |
Attribution & Acknowledgments
- Original Foundation: Jackrong/Qwopus3.8-27B-Flash & Qwen AI
- Abliteration Methodology: p-e-w/heretic
- 1M YaRN Scaling & Quantization Suite: Solstice-AI