During the weekend, the Chinese company of the Deepseek launched an AI chat application that includes a “reasoning” model comparable to OPENAI O1, causing a stir between American companies as Depseek rose to The top of Apple's app store.
Deepseek is a company based in Hangzhou, China, which provides generative models and the integration of AI. Its first products to make waves in the US market are the deep and R1 type GPT-4, an advanced “reasoning model”. Like ChatgPT, Deepseek-V3 and R1 quickly respond to indications in natural language.
The actions of Nvidia and Microsoft fell on Monday after the debut Buzzy. In general, the Stock Market reflected a sudden fall in confidence in the manufacturers of the US. UU. The success of Deepseek caused a conversation about whether US restrictions on Chinese access to the chips of ia limited or encouraged competition.
For technology professionals, Depseek offers another option to write code or improve efficiency around daily tasks. Together that Deepseek's R1 model is able to explain its reasoning, it is based on a family of open source models that can be accessed in Github.
What is notable of Deepseek?
Like the OPENAI O1 (previously known as Strawberry), the reasoning model slows its prediction capabilities to “reason” through your work, which helps you provide more precise answers. In particular, reasoning models have obtained well in reference points for mathematics and coding.
Deepseek said Depseek-V3 obtained a higher score than GPT-4O in the MMLU and Humaneval tests, two of a battery of evaluations that compare AI's responses.
Deepseek said that one of its models cost $ 5.6 million to train, a fraction of money often spent on similar projects in Silicon Valley.
You can access Deepseek-V3 and R1 through the App Store or in a browser. Visitors from the Deepseek site can select the R1 model for slower answers to more complex questions. When selected, the R1 model creates long answers that explain in a conversation style how it came to their conclusions.
Until Monday morning, the site warned to the Depseek chat site can be interrupted, although chatbot worked normally.
Deepseek also offers an APII, which works through the OpenAI SDK or software compatible with OpenAi SDK.
See: Operai announced operator, an AI agent who can take several steps actions in a web browser, such as choosing flights.
What does the launch of Deepseek's V3 and R1 mean for the AI industry?
“We can completely expect an application ecosystem in R1, as well as in several global cloud suppliers that offer their models as a consumable API,” said the vice president of Gartner, Arun Chandrasekaran, in an email to Techrepublic. “Depseek's future success is based on its ability to innovate continuously (instead of being a unique success), building an ecosystem of developers in their products and overcoming cultural barriers, given their country of origin.”
Chandrasekaran said that low cost, efficiency, Deepseek reference results and open weights make it remarkable.
Deepseek-V3 was trained in 2,048 GPU NVIDIA H800. American manufacturers are not, according to the export rules established by the Biden Administration, which are allowed to sell high -performance AI training chips to China companies.
“The potential power and low cost development of Deepseek are questioning the hundreds of billions of dollars committed in the United States,” said Ivan Feinaseth, a Financial Tigress Market Analyst, according to a note for customers acquired by ABC News .
Deepseek differs even more being an open source project and promoted by research, while OpenAi focuses more and more on commercial efforts.
“Deepseek R1 is one of the most surprising and impressive advances I have seen, and as an open source, a deep gift for the world”, “, the capitalist of Silicon Valley adventure, Marc Andreessen, published on X on Friday.
Gartner said that the world semiconductor industry of AI will reach $ 114,048 in 2025. Gartner predicted that the power required for the data centers to execute the newly added the IA servers will reach 500 Terawatt-Horas by 2027.
Depseek presents multimodal models
On Monday, Depseek continued its success with another surprise: the Janus-Pro Family of multimodal models. These models can analyze and generate images.
