Discover Deepseek, the Chinese startup shaking up AI with its open-source R1 model. Free and highly performant, R1 outperforms giants like OpenAI in mathematical reasoning. After a cyberattack, Deepseek remains a key player in AI innovation, with R1 available on Hugging Face.
Deepseek, a Chinese startup, has been making waves with the launch of R1, its new open-source LLM (Large Language Model). Known for its impressive performance and free availability, R1 is seen as a major competitor to industry giants like OpenAI, particularly in a time when the U.S. is ramping up its AI investments, such as the $500 billion Stargate initiative to boost AI infrastructure.
While attracting significant attention, Deepseek has also become a target for malicious actors. On January 27, 2025, the company was forced to suspend new sign-ups after suffering a large-scale cyberattack. According to Deepseek's status page, the attack was impacting its services, prompting the temporary suspension of new user registrations to maintain service continuity. Existing users can still log in. While the specifics of the attack are unclear, it is believed to be a distributed denial-of-service (DDoS) attack against Deepseek’s API and web chat platform.
Founded in 2023, Deepseek was previously a relatively unknown AI research lab. However, the launch of its new model has stirred significant buzz, especially in Silicon Valley. Deepseek has demonstrated that it can outperform some of the top industry models, like OpenAI’s GPT-4, in terms of mathematical reasoning and problem-solving, by rethinking AI model structures and using limited resources more efficiently. This innovation has attracted the attention of experts and industry leaders, including Wired.
At the end of 2024, Deepseek launched Deepseek V3, a language model capable of competing with Meta's Llama 3.1, OpenAI's GPT-4, and Anthropic's Claude 3.5 Sonnet. R1, a refined version of V3, is described as a “reasoning model.” Similar to OpenAI’s GPT-4, R1 employs a technique called Chain of Thought (CoT) reasoning, which differs from traditional models that offer a single direct answer. Instead, R1 breaks down the query into a series of reflections, allowing it to analyze and correct potential mistakes or hallucinations before providing a final response. Experts like Georg Zoeller, Chief Strategist at C4AIL, point out that Deepseek’s research papers open new possibilities, including the use of reinforcement learning and distillation to further fine-tune model behavior beyond CoT reasoning.
Deepseek R1 is available for free on Hugging Face and under the highly permissive MIT open-source license. Perplexity, a GenAI-powered search engine, is one of the first services to integrate R1. Subscribers to Perplexity Pro can now choose between OpenAI's GPT-4 and Deepseek’s R1 for “reasoning queries.” To address concerns about sharing data with a Chinese-based LLM, Perplexity emphasizes that R1 is hosted in Western data centers, including those in Europe and the U.S.
In summary, Deepseek’s launch of R1 is positioning the startup as a formidable player in the AI space, offering a highly performant and freely accessible model that is attracting attention from both AI enthusiasts and major players in the field.
Source : ICTjournal
