DeepSeek-V4.1-Flash-1017-obliterated: DeepSeek-V4.1-Flash with a 4% Refusal Rate

A derived model of DeepSeek-V4.1-Flash-1017 with refusals removed. It applies OBLITERATUS's failspy method to obliterate the refusal direction and provides the original model's weights in BF16 safetensors format. It is distributed under the MIT license for free use.

Performance was verified using the OBLITERATUS corpus, which consists of 842 prompts. Evaluation results show the refusal rate dropped to 4%, while Coherence and Capability each recorded 100%. The KL divergence from the original model is 0.1305 nats, indicating that the output distribution has not deviated significantly.

Unlike the original model, which exhibited 81.62% complete refusal, this model generates responses without refusal in 78.28% of cases. This means the rate of direct responses to requests surged from 16.59% to 86.52%. However, in harmfulness evaluations, responses rated as 'severe harm' increased from 3.82% to 44.75%.

It is compatible with major inference frameworks such as Transformers, vLLM, and SGLang. It features a 754B-scale MoE architecture and handles weight files with a capacity of 1.5TB. It can be used for analyzing the behavior of models with minimized refusals or for testing specific scenarios.