DeepSeek paper offers new details on how it used 2,048 Nvidia chips to take on OpenAI
artificial intelligence (AI) research lab
DeepSeek has released a new research paper revealing in detail for the first time how it built one of the world's most powerful open-source AI systems at a fraction of the cost of its competitors.
'Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures', co-authored by DeepSeek founder Liang Wenfeng and released on Wednesday, attributes the start-up's breakthrough in training high-performance, cost-efficient AI systems to a hardware-software co-design approach.
'DeepSeek-V3, trained on 2,048 Nvidia H800 GPUs, demonstrates how hardware-aware model co-design can effectively address these challenges, enabling cost-efficient training and inference at scale,' the researchers wrote. DeepSeek and its hedge fund owner High-Flyer had previously stockpiled the H800, which
Nvidia originally designed for the China market to comply with US export restrictions but were banned from export to to the country in 2023.
The start-up's training approach stemmed from the team's awareness of hardware constraints and the 'exorbitant costs' of training large language models (LLMs) – the technology behind AI chatbots such as OpenAI's
ChatGPT – according to the paper.
The paper details technical optimisations that boost memory efficiency, streamline inter-chip communication, and enhance overall AI infrastructure performance – key advancements for reducing operational costs while scaling capabilities. These offer a 'practical blueprint for innovation in next-generation AI systems', the researchers said.
Play
DeepSeek also highlighted its use of a mixture-of-experts (MoE) model architecture, a machine-learning approach that divides an AI model into separate sub-networks, or experts, each focused on a subset of the input data while working collaboratively.
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
South China Morning Post
3 hours ago
- South China Morning Post
DeepSeek's namesake chatbot sees a drop in downloads as AI apps for work, education rise
DeepSeek 's namesake chatbot recorded lower downloads and user numbers in the second quarter, according to a new report, as artificial intelligence (AI) apps specifically designed for study or office work gained traction. Advertisement In the three months to June, the average monthly download of DeepSeek's chatbot fell 72 per cent to 22.6 million from the previous quarter, when the Hangzhou -based start-up benefited from the wide popularity of its V3 and R1 AI models, according to Monday's report by market research firm QuestMobile in collaboration with state-funded newspaper National Business Daily. While the chatbot's monthly active users (MAUs) of 170 million still topped the mainland Chinese market, that number was down 9 per cent quarter on quarter. DeepSeek's chatbot was overtaken by ByteDance -owned Doubao , which posted average monthly downloads of 29.8 million in the second quarter, up 9.5 per cent from the previous three-month period, the QuestMobile report showed. Doubao's MAUs grew 30 per cent to 130 million in the June quarter. Some other popular general-purpose chatbots also saw their downloads decline in the second quarter, which reflected a shift in which consumer-facing AI apps are now more popular in China. Advertisement The average monthly downloads for Tencent Holdings Yuanbao tumbled 54 per cent to 61.8 million, while those of Moonshot AI 's Kimi fell 58 per cent to 35.3 million, according to QuestMobile. 'This trend clearly shows that the phase of mere 'chatbots' is over,' the report said.
South China Morning Post
5 hours ago
- South China Morning Post
Alibaba upgrades flagship Qwen3 model to outperform OpenAI, DeepSeek in maths, coding
Alibaba Group Holding unveiled an upgraded version of its third-generation Qwen3 family of large language models (LLMs), improving one of its members to score higher in maths and coding than products from rivals OpenAI and DeepSeek. The new Qwen3-235B-A22B-Instruct-2507-FP8 is an open-source model that achieved 'significant improvements in general capabilities, including instruction following, logical reasoning, text comprehension, mathematics, science, coding and tool usage', according to a Tuesday update on artificial intelligence (AI) community HuggingFace and ModelScope, Alibaba's open-source platform. Alibaba owns the Post. It outperformed some rivals in certain assessments, such as the 2025 American Invitational Mathematics Examination, where the new Alibaba model scored 70.3. By comparison, DeepSeek-V3-0324, the most recent version of the foundational model that was released in March, scored 46.6 while OpenAI's GPT-4o-0327 scored 26.7. As for coding capabilities, the new Qwen secured 87.9 points from the MultiPL-E benchmark, slightly higher than 82.2 and 82.7 from the DeepSeek and OpenAI models above, respectively, though it lagged behind Claude Opus 4 Non-thinking, from Anthropic, which scored 88.5. Alibaba's new release was an upgrade from the Qwen3-235B-A22B-FP8. But it only supports non-thinking mode, where an AI system provides a direct output without the explicit reasoning steps or chain of thought that a thinking model might employ. As a result, its content length was boosted eightfold to 256,000 tokens, making it able to handle longer texts in a single conversation. Also on Tuesday, Alibaba said a Qwen model with 3 billion parameters would be integrated into HP's smart assistant 'Xiaowei Hui' on its personal computers in China, enhancing capabilities including drafting documents and summarising meetings.
South China Morning Post
11 hours ago
- South China Morning Post
The wisdom of Australia's fresh approach to China
Feel strongly about these letters, or any other aspects of the news? Share your views by emailing us your Letter to the Editor at letters@ or filling in this Google form . Submissions should not exceed 400 words, and must include your full name and address, plus a phone number for verification Australian Prime Minister Anthony Albanese's recently concluded visit ('How deals are trumping port dispute on Australian PM Albanese's China visit', July 17 ) is a welcome breath of fresh air in Canberra's approach to China. Since the early 2000s, Australia's economic prosperity has been closely linked with China – Australians became wealthy selling China iron ore, coal and other natural resources that helped power China's extraordinary economic growth and societal transformation. Under the previous Liberal-National coalition government, Australia took an unwise turn in its foreign policy by uncritically siding with the United States and needlessly antagonising China – which culminated in Canberra effectively blaming China for Covid-19 by calling for an independent inquiry into its origins. The past year has shown a world undergoing seismic changes. It is becoming clear that China will emerge as a dominant, if not the dominant, country that masters the industries of the future – such as renewable energy, electric vehicles, biotechnology and artificial intelligence, as most vividly demonstrated by the splashy emergence of DeepSeek's AI model to public awareness.
