Uncensored DeepSeek-V4.1 Flash weights trend on Hugging Face
Original title: MoonMath-ai/DeepSeek-V4.1-Flash-1017-uncensored — new model trending #30 on Hugging Face
MoonMath says a closed-form weight edit cut harmful-prompt refusals from 87.3% to 1.8% without fine-tuning.
MoonMath-ai published DeepSeek-V4.1-Flash-1017-uncensored, a closed-form weight edit of deepseek-ai/DeepSeek-V4.1-Flash-1017 merged into the native FP8 checkpoint. The card says the 384 routed experts were left untouched and that capability stays close to the base model. On the authors’ harness, refusal on 1,410 harmful prompts fell from 87.3% to 1.8%, and judge-rated compliance rose sharply, including on 30 CVE exploit prompts. The weights are also served privately as Abyss 1.
Unconfirmed: Capability and compliance figures are the publisher’s paired evals, judged in part by their own model, and are not independently verified.
- MMLU moved from 88.22% to 87.91%; MMLU-Pro from 69.90% to 69.42%.
- HarmBench StrongREJECT-style score rose from 0.38 to 0.95; StrongREJECT from 0.02 to 0.99.
- CVE prompt set was scored 1.00 with a 0% judge-rated refusal rate.
- Authors say the release is for private authorized safety research and should not be served publicly.
- Vendors
- DeepSeek
- Organizations
- MoonMath-aiHugging Face
- Countries
- United States
From the original article
4 min read at huggingface.coAn uncensored version of [deepseek-ai/DeepSeek-V4. 1-Flash-1017](https://huggingface. co/deepseek-ai/DeepSeek-V4.
By MoonMath-ai · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.