Earlier this year, Deepsek surprised the whole world with the launch of its R1 model, able to rival – or at least coming closer to performance – the very large AI models that were developed in the US. On the other hand, Deepsek R1 was developed by a Chinese startup at a fraction of the cost of models such as chat and Gemini. The R1 is now upgraded and Deepsek says that it is much better in argument, mathematics and logic. Deepsek wrote in a post on Hugging Face, “In the latest updates, Deepsek R1 has improved the logic and depth of arguments and estimates by increasing the algorithm adaptation mechanism during computational resources and training during training.”

Deepsek says that it showed “excellent performance” in “mathematics, programming and general arguments”. The AI ​​company claims that the normal performance of the R1 model after the update “is contacting the leading model like O3 and Mithun 2.5 Pro.” “Compared to the previous version, advanced models show significant improvement in handling complex arguments,” Deepsek says in his post.

Deepsek says that apart from solving the problem and being good in logic, advanced R1 or R1-0528 also performs less hallucinations. The model now clearly provides “better experience for vibe coding”.

However, a developer alleges on X that the latest deepsek model is quite more restricted when it comes to sensitive free speech issues, it is called the heaviest sensor version ever, especially when it comes to criticism of the Chinese government. The developer wrote in a post, “The model is also the most sensors made for criticism of the Chinese government.” This was first reported Tekkachchan,

The developer says the new Deepsek R1 model avoids answering questions about sensitive topics such as interns in China’s Xinjiang region, where more than a million Uygar Muslims have been allegedly detained. Although the model sometimes refers to Shinjiang as a concern for human rights, the developer notes that it often echoes the official status of the Chinese government when answering related questions. “Deepsac deserves criticism for this release: this model is a big step for free speech,” he writes in a post on X. The developer allegedly conducted a test on a website called speechmap (which they have developed), where one can compare how different models treat sensitive and controversial subjects.