Perplexity open-sources decensored DeepSeek R1 variant

Perplexity released R1 1776, its own post-trained version of DeepSeek’s R1 reasoning model, modified to answer questions on topics the original model deflected or refused following Chinese government content restrictions — such as Tiananmen Square, Taiwan and Xinjiang. Perplexity said its team identified roughly 300 topics subject to this kind of censorship, built a multilingual classifier to detect prompts touching them, and curated a dataset of about 40,000 examples for further post-training.

Perplexity reported the retrained model answered such prompts factually across a 1,000-example, multilingual evaluation set judged by both human annotators and LLM-based judges, where the original R1 had declined or deflected on many. It also reported that mathematics and knowledge benchmarks — including AIME 2024, MMLU, MATH-500 and DROP — were nearly unchanged from the base model, arguing that removing the censorship behaviour had come at negligible cost to reasoning ability. The retrained weights, including a smaller distilled Llama 70B version, were released under the same permissive MIT licence DeepSeek had used for R1 itself, and made available through Hugging Face and Perplexity’s own API.

The release illustrated a pattern that followed DeepSeek R1’s January launch: because R1’s weights and training method were open, other organisations could inspect and retrain around its behaviour rather than simply adopting or rejecting it, turning the state-alignment built into a Chinese model’s outputs into a problem other labs treated as separately fixable.