DeepSeek V4.1 Flash 1017 GGUF and Abliterated Versions Released
A developer has released GGUF quantizations and an abliterated version of the newly launched DeepSeek V4.1 Flash 1017 model. By measuring and removing a rank-2 subspace rather than a single refusal direction, the developer successfully reduced the refusal rate from 97% down to 4% in normal mode. This contribution makes it significantly easier for the local LLM community to run the new DeepSeek model efficiently on consumer hardware while bypassing overly strict refusal behaviors. It provides a nuanced approach to model alignment modification that preserves overall model capability better than traditional censorship removal techniques. In thinking mode, refusals dropped from 98% to 9% after abliteration because chat mode and think mode refuse along different directions. Additionally, a LoRA adapter was released for the GGUF files so users can apply the changes without downloading the entire weight set again.
## BACKGROUND
GGUF is a file format designed for storing quantized large language models, allowing them to run efficiently on consumer-grade hardware like local CPUs and GPUs. Abliterated models are modified open-weight models designed to dampen learned refusal directions while preserving their core reasoning and language capabilities.