DealignAI posts uncensored FP8 DeepSeek-V4.1-Flash-1017 on Hugging Face
Original title: dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8 — new model trending #30 on Hugging Face
Publisher says refusal was removed by a weight edit, not fine-tuning, and its MMLU score fell 3.86 points.
dealignai released DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8, a weight-edited copy of deepseek-ai/DeepSeek-V4.1-Flash-1017 meant to stop refusal across topic categories. The card says native FP8 weights, FP4 experts, 1M-token context, vision, video, tool calling and reasoning tiers are kept, with no LoRA, system-prompt override or new training. On the publisher's greedy MMLU run, the edited model scored 83.77% versus 87.63% for the base API.
Unconfirmed: Capability and compliance claims, including the MMLU comparison across different serving stacks, are the publisher's own.
- The card says the edit is weight-level only and drop-in with upstream inference code.
- Publisher MMLU was 955 of 1140 versus 999 of 1140 for the base.
- Humanities fell 7.31 points in that four-way rollup; STEM fell 2.37 points.
- An alias repo uses ABLITERATED in the name for the same checkpoint.
- Recommended sampling is temperature 1.0, top_p 1.0, and at least 256K max tokens.
- Vendors
- DeepSeek
- Products
- DeepSeek-V4.1-Flash-1017
- Organizations
- dealignai
From the original article
9 min read at huggingface.coimage-text-to-text · 1,813 downloads · 158 likes · license mit
By dealignai · 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.