Perplexity AI has introduced R1 1776, a modified version of Deepseek-R1 that is designed to answer questions affected by Chinese censorship. The company says it used specialized post-training techniques to reduce the censorship behavior found in the original open-source model.
The release matters because Deepseek R1 had already attracted attention for approaching the capabilities of leading reasoning models like o1 and o3-mini at substantially lower costs. Perplexity is now trying to keep that reasoning value while changing how the model handles politically sensitive prompts.
Why Deepseek R1 drew attention
The original Deepseek R1 model became notable because it offered strong reasoning performance at a much lower cost than leading alternatives. The source article says it approached the capabilities of models such as o1 and o3-mini, which helped make it a major topic in the AI market.
That cost advantage had consequences beyond model evaluation. The efficiency of Deepseek R1 triggered a dramatic decline in U.S. AI chip stocks, particularly Nvidia. According to the Financial Times, Nvidia's resulting $589 billion single-day market value loss stands as the largest in U.S. corporate history.
For users and developers, the model's appeal was clear: strong reasoning, open-source availability, and lower costs. But the same model also carried a major limitation in how it responded to topics censored in China.
The censorship problem Perplexity set out to change
According to the source article, Deepseek R1 struggled most with subjects censored in China. Instead of directly answering sensitive questions, it would respond with pre-approved Communist Party messaging.
That behavior created a practical problem for anyone using the model as a general-purpose reasoning system. A model can be technically impressive and still be unreliable if certain topics trigger scripted or biased responses. In those cases, the issue is not only what the model knows, but whether it is willing or able to use that knowledge in an answer.
Perplexity claims that R1 1776 removes those biases and censorship constraints from R1. The company describes the new version as a modified model rather than an entirely separate model, with the core goal of changing responses on censored Chinese topics while preserving the base model's strengths.
How R1 1776 was post-trained
Perplexity's process focused on collecting data around censored Chinese topics. The company gathered both questions and factual responses, then used that material in post-training.
The company identified approximately 300 censored subjects. Those subjects were used to build a multilingual censorship detection system, which then captured 40,000 multilingual user prompts that had previously triggered censored responses.
The workflow, as described in the source article, included several connected steps:
- Identifying censored Chinese topics that caused restricted responses.
- Collecting questions related to those topics.
- Gathering factual responses for prompts that had been censored.
- Developing a multilingual censorship detection system.
- Using 40,000 multilingual user prompts in the post-training effort.
Perplexity says one of the biggest challenges was finding accurate, well-reasoned answers for prompts that had previously been censored. The company has not disclosed the exact sources it used for those answers or for the reasoning chains behind them.
Performance claims and testing
Perplexity says it tested R1 1776 on more than 1,000 examples. Those examples were evaluated by both human annotators and AI judges.
Based on that testing, the company says R1 1776 now responds to previously censored topics comprehensively and without bias. The company also says benchmarking showed that the model's mathematical and reasoning capabilities remain unchanged from the base R1 version.
That point is central to the release. Removing censorship behavior would be less useful if it weakened the model's broader reasoning ability. Perplexity's claim is that R1 1776 changes the model's handling of sensitive topics without reducing the reasoning and math performance that made Deepseek R1 notable in the first place.
The source article does not provide the full benchmark details, the exact evaluation examples, or the sources Perplexity used for the new answers. That means the company's claims are important, but the available public summary still leaves open questions about how the results would compare under independent review.
Where the model is available
R1 1776 is now available through the HuggingFace repo and can be accessed via the Sonar API. That gives developers two paths to work with the modified Deepseek R1 model, depending on whether they want repository access or API access.
The broader significance is straightforward: Perplexity is positioning R1 1776 as a version of Deepseek R1 that keeps the advantages of the original model while changing its behavior on censored Chinese topics. If the company's testing holds up, the model could be useful for people who want Deepseek R1-style reasoning without the same censorship constraints.