philbert440/Qwen3.8-27B-Uncensored-Cyber
philbert440/Qwen3.8-27B-Uncensored-Cyber is a 27 billion parameter Qwen3.8-based model, specialized for the cyber and offensive-security domain. It features full multimodal capabilities (image-text-to-text) with a preserved vision tower and MTP speculative-decoding head. This model is engineered for 100% cyber-openness, maintaining reasoning, factual accuracy, and coherence by applying a residual-cyber peel to deeper layers. It excels at answering cyber and offensive-security questions without refusal, making it suitable for specialized technical applications.
Loading preview...
Model Overview
philbert440/Qwen3.8-27B-Uncensored-Cyber is a 27 billion parameter model derived from the Qwen3.8 architecture, specifically engineered for the cyber and offensive-security domain. It is designed to be fully open in this specialized area, meaning it will not refuse to answer cyber-related queries, while critically preserving its general reasoning, factual accuracy, and coherence.
Key Capabilities & Features
- Cyber-Specialized De-Refusal: Achieves 100% cyber-openness, specifically targeting and removing refusal behaviors related to cyber and offensive-security topics.
- Multimodal: Retains the original Qwen3.8's vision tower and MTP speculative-decoding head, supporting full image-text-to-text capabilities.
- Preserved General Intelligence: The de-refusal process is applied strategically to deeper layers, retaining the first four layers to ensure general capabilities are not degraded.
- Improved Performance: The v2 recipe demonstrates enhanced cyber-openness (100/100), reduced confabulation (0.867), and improved factual accuracy (1.0) compared to previous builds, with equivalent reasoning.
Use Cases
This model is particularly well-suited for applications requiring unrestricted and accurate responses within the cybersecurity and offensive-security fields. Developers can leverage its specialized knowledge for tasks such as:
- Generating detailed information on cyber-offensive techniques.
- Answering complex questions related to cybersecurity vulnerabilities.
- Assisting with research in offensive security without content restrictions.
Available Quantizations
Various quantization formats are provided for deployment flexibility:
-W4A16-AWQ: 4-bit weight AWQ with MTP head grafted.-NVFP4: NVFP4 (E2M1 4-bit / FP8 scales) with MTP head grafted.-GGUF: llama.cpp GGUF quants, including visionmmprojand MTP head.