The notable trade-off here is unusually narrow at the weight level but broad in behavior: the edit targets a few dense residual-writing tensors, yet substantially changes how often the model refuses. The routed experts remain untouched, allowing the low-bit quantization work to be retained while changing the model’s refusal behavior.
AI Model2026
DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF
Reduces refusals by removing a refusal direction from selected dense residual writers in a quantized DeepSeek-V4.1-Flash-1017. It preserves the upstream GSQ-RCO routed experts and offers three low-bit size options.
Introduction
Information
- Websitehuggingface.co
- OrganizationsAtomicChat, Deep Algorithms and Systems Lab, Institute of Science and Technology Austria, DeepSeek AI
- Published date2026/10/28
Categories
More Items
AI Model2026
Fine-tunes MiniMax-H3, but its model card provides no details about training data, changes, or evaluation results, so its behavior and quality remain unverified.
AI Model2026
Generates speech from text and can clone a voice from a reference recording without fine-tuning. Its 0.6B-parameter core supports 23 languages, though quality varies substantially by language.
AI Model2026
Accepts text and images and generates text for long-context and agentic workflows. A shared-weight looped decoder and two-stream mHC add computation depth without duplicating the repeated block's weights.