DeepSeek has launched an advanced version of its large language model, DeepSeek-V3-0324, narrowing the gap with AI giants OpenAI and Anthropic. The latest model, now available on Hugging Face, has garnered attention for its enhanced reasoning and coding abilities.
According to DeepSeek, the new iteration outperforms its predecessor across multiple benchmarks, particularly in mathematical problem-solving and software development. The model has demonstrated significant improvements, scoring 59.4 on the American Invitational Mathematics Examination (AIME), a leap from the previous 39.6.
On LiveCodeBench, a coding assessment, it gained 10 points, reaching 49.2. With 685 billion parameters, the upgraded version slightly surpasses its predecessor’s 671 billion. DeepSeek has made the model more accessible by shifting from a proprietary license to an MIT license, allowing developers worldwide to utilize it freely.
Experts have commended the model’s capabilities, with Kuittinen Petri, a lecturer at Häme University of Applied Sciences, remarking that “Anthropic and OpenAI are in trouble.” Petri tested the model by prompting it to create a responsive website for an AI company, and it successfully generated a fully functional, mobile-friendly page with 958 lines of code.
Apple research scientist Awni Hannun also evaluated the model’s performance on a 512GB M3 Ultra workstation, noting its efficiency in processing over 20 tokens per second while managing memory consumption effectively.
DeepSeek’s rapid progress in AI development has sparked speculation about future releases. After launching its V3 model in December and the R1 model in January, discussions are underway about an upcoming R2 version focused on advanced reasoning.
AI professionals, including Jasper Zhang and Fahd Mirza, have lauded DeepSeek-V3-0324’s problem-solving skills, with Mirza describing its performance as “mind-blowing.” Despite operating with significantly fewer financial resources than its competitors, DeepSeek continues to make remarkable strides in AI innovation.