DeepSeek-R1 AI model

The main articles are:

• Neural networks (neural networks)
• Speech Recognition (Technology, Market)
• Speech technology: On the path from recognition to understanding

2025: Version of the "reasoning" AI model, superior in performance to OpenAI's o1

The Chinese company DeepSeek has released a model of artificial intelligence DeepSeek-R1, which, according to the developers, demonstrates performance comparable to the o1 model from OpenAI. At the same time, its code is open, and it also works in Russia without restrictions.

1 million output tokens in the DeepSeek AI model cost $2.19 dollars compared to $60 for the model ChatGPT from. OpenAI Nvidia Stocks collapsed. Nvidia's triple-leverage traders took a one-day 52% loss.

Nvidia Corp said the DeepSeek AI model is an "excellent AI achievement" that meets the requirements of controlling the export of technology to the United States.

DeepSeek-R1 is capable of self-testing, reflection, and generating long chains of reasoning. The company shared the model's results in various tests evaluating AI abilities, Hi-Tech Mail reported on January 20, 2025, citing DeepSeek.

The evaluation of the mathematical abilities of the DeepSeek-R1 was carried out on two different benchmarks: MATH-500 and AIME 2024. In the first, the model reached 97.3%, slightly more than the o1 from OpenAI (96.4%). In the second DeepSeek-R1, he scored 79.8%, and o1 - 79.2%. In the test for logical thinking and general knowledge (MMLU), DeepSeek-R1 showed a result of 90.8%, which is close to the OpenAI-o1-1217 indicator (91.8%).

𝓲

Фото: DeepSeek

Assessment of DeepSeek-R1 in tasks requiring programming skills was carried out using SWE-bench Verified, Codeforces and GPQA Diamond benchmarks. In the first DeepSeek-R1, the result is 49.2%. This figure is slightly higher than that of o1 (48.9%). On the Codeforces platform, the Chinese model reached 96.3%, just below the o1 result (96.6%). In the GPQA Diamond test, the DeepSeek-R1 result was 71.5% and o1 - 75.7%. At the same time, R1 bypassed o1-mini in all benchmarks.

The R1 is comparable in performance to the OpenAI model, but unlike it is fully open and available for free use and commercialization under the MIT license. You can evaluate the possibilities of the DeepSeek-R1 for free. Source code is available for developers on GitHub.

The DeepSeek neural network first appeared in 2023, and in two years the AI model has been updated three times. Before R1, users had access to the DeepSeek V3 version, which was called "one of the most powerful on the market." In some tests, the DeepSeek V3 performed higher than the Llama 3.1 and OpenAI GPT-4o models. In mid-January 2025, the developers of the chatbot released an official mobile application for Android and iOS - it can also be downloaded in Russia^[1].

Notes