DeepSeek has launched an upgraded version of its large language model, DeepSeek-V3-0324, closing gap with OpenAI and Anthropic.
The new model is now available on Hugging Face, an AI development platform, where it has already gained attention for its improved reasoning and coding capabilities.
DeepSeek claims that this version surpasses the previous model in multiple benchmarks, particularly in mathematical problem-solving and software development.
One of the improvements is its performance on the American Invitational Mathematics Examination (AIME), where it scored 59.4, a notable jump from the previous model’s 39.6.
Similarly, on LiveCodeBench, a coding assessment, it gained 10 points to reach 49.2. These improvements suggest a more capable AI system for both research and practical applications.
With 685 billion parameters, DeepSeek-V3-0324 slightly surpasses the earlier V3 model’s 671 billion. Unlike old models, which used a proprietary commercial license, the latest version is distributed under the MIT license, making it more accessible to developers worldwide.
DeepSeek’s development has caught the attention of experts. “Anthropic and OpenAI are in trouble,” said Kuittinen Petri, a lecturer at Häme University of Applied Sciences, on X.
He tested the model by instructing it to “create a great-looking responsive front page for an AI company,” and it successfully generated a fully functional, mobile-friendly website with 958 lines of code.
Apple research scientist Awni Hannun ran the model on a 512GB M3 Ultra workstation. While it processed over 20 tokens per second, he noted that memory usage peaked at 381GB, which, though high, remained within expectations for a model of this scale.
DeepSeek’s progress in AI development has been commendably fast. After launching its V3 model in December and the R1 model in January, speculation is already building about the release of R2, a possible follow-up to its reasoning-focused series. “The coding capabilities are much stronger, and the new version may pave the way for the launch of R2,” said Li Bangzhu, founder of AIcpb.com.
Jasper Zhang, a University of California, Berkeley graduate and maths Olympiad gold medallist, also tested the model using an AIME 2025 problem. “It solved it smoothly,” he noted, asserting that DeepSeek’s models are closing the gap with their Western competitors.
Fahd Mirza, lead cloud and AI engineer at Australian construction materials company Boral, described DeepSeek-V3-0324 as “mind-blowing.” On his YouTube channel, he shared a demonstration of the model tackling complex coding and mathematical tasks, calling its performance “outstanding.”
DeepSeek’s approach to AI development has focused on efficiency and accessibility. Unlike heavily funded rivals, it operates with significantly fewer financial resources. Petri pointed out, “DeepSeek is doing all this with just [roughly] 2 per cent [of the] money resources of OpenAI.”
