DeepSeek quietly rolled out an updated AI model

      While major American IT companies are competing in the number of zeros in their AI development budgets, China's DeepSeek has decided to surprise everyone again. Without presentations, fanfare, or Elon Musk, she just posted a new version of her DeepSeek R1 reasoning model on Hugging Face. Quiet, calm, in Chinese.

       DeepSeek made a lot of noise at the beginning of 2025 — their first version of R1, let me remind you, turned out to be not just good, but really powerful: it surpassed models from Meta* and even OpenAI in logic and reasoning. At the same time, this development cost suspiciously little and was done suspiciously quickly. There was a panic on Wall Street then - the shares of Nvidia and other AI giants went down sharply. Well, DeepSeek just went into the shadows... to return.

      R1 is back in business, and now even smarter

       The new version, called R1-0528, was released again without any press releases. But it was the experts who noticed. On LiveCodeBench, the model occupies an honorable third place, second only to the o4-mini and o3. Even Grok 3 from xAI and Chinese Qwen 3 from Alibaba remained behind.

       The first users are already saying that the model has become "sharper in logic, stronger in mathematics and code," and also almost does not carry nonsense — its level of "hallucinations" is much lower. That is, now she is not only smart, but also less inclined to invent facts.

       Let's see how the new version of DeepSeek R1 can be useful not to a technician, but to a businessman who thinks in the categories of "profit", "cost reduction" and "process acceleration" rather than "ML benchmarks".

      

      

       R1 is able to reason, build logical chains, understand data and explain what's what. For example, if you have an Excel spreadsheet with sales results for the year, you can ask: "Why did sales drop in March?" and get not just numbers, but a sane, logical answer with hypotheses.

      

      

      

      

       Do you need to make a beautiful report for investors? Prepare a presentation for the client with figures, forecasts and conclusions? DeepSeek can handle it. He writes logically, clearly, and — most importantly — to the point, without inventing nonsense, as is often the case with other models.

      

      

      

      

       Set tasks, write standard letters, generate documentation, prepare the text of the contract — all this can be entrusted to R1. It is especially good where logical thinking is needed, not just text generation. For example: "Write a delivery contract, but keep in mind that the client is a foreigner and pays with a delay."

      

      

      

      

       Do you need to quickly test a hypothesis, build an MVP (minimum viable product) or a bot in Telegram? DeepSeek can handle the code as well. And no worse than expensive closed solutions. And you don't have to pay $20-30 per month, as you can use GPT or Gemini — R1 locally or via the cloud, and pay pennies for hosting.

      

      

      

      

       The model is open source and distributed under the MIT license. This means that it can be integrated into any project, used in commerce, and modified without fear of courts and restrictions.

      

      

      The smart one won't go uphill. China is smart

      Interestingly, the situation with DeepSeek affects the entire processor manufacturing industry. The neural network proves by its existence that China doesn't really need top-end American chips. While the United States is trying to limit access to semiconductors, Chinese developers are betting on software optimization. And they do it very well.

       Nvidia CEO Jensen Huang has already stated: "China already has AI. The question is not whether it will be there — it already is." And, judging by DeepSeek, he's not just "there," he's also on the heels of the Western giants.

       An additional bonus is that the model is open, free, and licensed under MIT. If you want to use it in commerce, you're welcome. If you want to finish it, it's not a question. It's nice when technology is not only powerful, but also not greedy.

       Meta is recognized as an extremist organization, and its activities in the territory of the Russian Federation are prohibited.

       Meanwhile, Elon Musk is trying to send millions of robots in factories, Tesla, and Google is teaching its cars to think "inner voice", "Yandex" has decided to participate in the global race of humanoids in their own way. The company will focus on software that will allow robots not only to move around, but also to feel quite comfortable in the office, in the warehouse and even at your home.