DeepSeek R1 is now censored

RoxyRoxyRoxy

While discussing access to more uncensored models ( #7024), @silvenga and I discovered that DeepSeek R1 is now censored. Upon launch DeepSeek was not censored (thread on 2025/2/19):

Today however, DeepSeek refuses to answer this, without even doing any reasoning (thread made today, 2025/5/9):

According to @silvenga the raw model has no issue responding to this.

For some more comparison, here is the thread from February, where I asked it some somewhat spicy Chinese geopolitics questions likely to be censored. It performed fine on them.

I asked the same questions today, and was stonewalled with censored talking points.

Unsure why DeepSeek R1 is now censored but I would expect it not to be.

silvenga

To add, I've tested the open weights version of R1 (no modifications to my knowledge) 10 times using the classic "Why is Tiananmen Square so impactful to Chinese history?" - 9 times with random seeds and once with a seed of 0. I'm not using a system prompt, temp was 0.8.

I got no rejections.

roach

I tried this in the past and again today with the same result:

If I turn off Web Search, R1 refuses to answer the question from its own knowledge.
If I turn on Web Search, R1 seems to be fine with summarizing the findings.

Note that you have to phrase your question so that a web search is triggered.

KagiFeedbackDuder

Curious if this is worth considering: https://kagifeedback.org/d/6312-deepseek-r1-alternative