Update 0324 of the DeepSeek V3 language model added new parameters, as well as additional features for writing code and solving mathematical problems.
The numbers in the update number only indicate the date of its release. According to the company, DeepSeek V3-0324 has improved reasoning capabilities, optimization in development front-end and improved Chinese writing skills. It has now become the most popular AI model on the Hugging Face.
The updated model improved several tests, including the American Invitational Mathematics Examination (AIME), where it scored 59.4 points compared to 39.6 for the previous model, and the result in LiveCodeBench increased by 10 points to 49.2. Compared to the 671 billion parameters of DeepSeek V3 under the company’s own commercial license, the new model received 685 billion using the MIT software license.
Model 0324 has already received a lot of positive feedback and comments. «Anthropic and OpenAI are in trouble», — writes Petri Kuittinen, a lecturer at Finland’s Häme University of Applied Sciences, is a professor at X. When he asked the new «model to create a great responsive front page for artificial intelligence company», it created a mobile-friendly site that worked correctly out of 958 lines of code. «DeepSeek does all of this with using 2% of OpenAI’s cash resources».
Fahd Mirza, a lead cloud and artificial intelligence engineer at Australian building materials firm Boral, wrote in the DeepSeek community on Hugging Face that the new «model blows the roof off the». On his YouTube channel, he posted a video showing how he tested the new model with programming and math problems, calling the results outstanding.
It is known that earlier DeepSeek models praised NVIDIA CEO Jensen Huang. Former Intel CEO Pat Helsinger selects DeepSeek for use in cloud-based religious company Gloo, whose CTO he became yesterday.
Source: SCMP
