DeepSeek quietly updates R1 AI model amid anticipation for next-gen tech
artificial intelligence (AI) start-up
DeepSeek quietly released a new version of its R1 reasoning model on Wednesday, marking its first revision since its high-profile debut in January.
The Hangzhou-based company said it had 'completed a minor update to the R1 model', which is now available on the website for its namesake chatbot, as well as its mobile apps, according to a notice posted in a company-run WeChat group chat.
DeepSeek did not disclose details of the changes in the update, dubbed R1-0528, which is now live on the open-source AI platform Hugging Face.
DeepSeek did not immediately respond to a request for comment.
The company last updated its foundational large language model V3 in March, touting improvements in coding and writing in the V3-0324 release on Hugging Face.
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
South China Morning Post
2 days ago
- South China Morning Post
How China's new ‘Darwin Monkey' could shake up future of AI in world first
Chinese engineers have unveiled the world's first brain-like computer made up of more than 2 billion artificial neurons. The neuron count of the 'Darwin Monkey' approaches that of a macaque and could be used to advance human brain-inspired artificial intelligence (AI), according to its developers at Zhejiang University. The Darwin Monkey is the latest generation of brain-inspired computers produced by Zhejiang University researchers. 'This is the world's first brain-like computer based on a dedicated neuromorphic chip with more than 2 billion neurons,' the university said on its social media account on Saturday. The computing system, made up of 960 Darwin 3 brain-inspired computing chips creating over 100 billion synapses, is 'a step closer to achieving more advanced brain-like intelligence', it said in the post. The Darwin Monkey has been successfully deployed to complete tasks like content generation, logical reasoning and mathematics, using the groundbreaking Chinese AI company DeepSeek's brain-like large model.
South China Morning Post
5 days ago
- South China Morning Post
DeepSeek founder shares best paper award at top global AI research conference
A research paper co-authored by Liang Wenfeng, founder of Chinese artificial intelligence start-up DeepSeek, was honoured with the best paper award at the Association for Computational Linguistics (ACL) conference in Vienna, Austria, widely recognised as the premier global conference for AI researchers. The paper, titled 'Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention,' was published on February 27, with Liang listed as one of 15 authors. The 'native sparse attention' mechanism is a core improvement that underpins the high efficiency and low-cost performance of DeepSeek's AI models. The paper's win comes as Chinese scientists and researchers are outperforming US peers in basic research in the field of computational linguistics and natural language processing. At this year's ACL conference, more than half of the first-named authors on accepted papers originated from China, up from less than 30 per cent last year. The US ranked second, with 14 per cent of first-named authors, according to ACL data. Among the four best papers recognised by ACL, two author teams were from China. They included Liang's DeepSeek team and Yang Yaodong's team from Peking University. An undated photo of DeepSeek's Liang Wenfeng. Photo: Weibo Yang, an assistant professor at the Institute of Artificial Intelligence and chief scientist of the Peking University-PsiBot Joint Laboratory, led research that explored a possible mechanism explaining the fragility of alignment in language models, attributed to the elasticity of language models.
AllAfrica
7 days ago
- AllAfrica
China's AI leap elevating stealth fighter ambitions
The South China Morning Post (SCMP) has reported that Chinese scientists have developed advanced aircraft-design software they claim breaks the 'curse of dimensionality,' a computational barrier that contributed to the US Navy's cancellation of its X-47B stealth drone program in 2015. Led by Huang Jiangtao at the China Aerodynamics Research and Development Center, the team introduced a geometric sensitivity computation method that enables optimization of hundreds of variables—such as stealth, aerodynamics and propulsion—without increasing computational load. Unlike traditional methods that grow exponentially more complex, their approach decouples gradient computation costs from design intricacy and integrates radar-absorbent materials directly into aerodynamic sensitivity equations. Their paper, published in Acta Aeronautica et Astronautica Sinica, demonstrated dramatic improvements using the X-47B as a case study. The researchers say this breakthrough could provide critical technical support for next-generation low-observable aircraft, including China's J-36 and J-50 fighters and stealth drones. As sixth-generation fighter programs worldwide face delays or cancellations, China's approach—emphasizing algorithmic efficiency over raw computing power—may save time and resources in stealth warplane development. The SCMP has also previously reported that China's Shenyang Aircraft Design Institute is using the DeepSeek AI platform to tackle complex engineering challenges and reduce time spent on technical reviews, freeing researchers to focus on core innovation tasks. Lead designer Wang Yongqing has stated that the technology is already generating new ideas and approaches for aerospace development, and confirmed steady progress on new variants of the multi-role J-35 stealth fighter. This progress may be underpinned by China's development of increasingly capable AI models. Nature reported this month that Moonshot AI's Kimi K2, an open-weight agentic large language model, matches or surpasses Western and DeepSeek models. The report indicates that Kimi K2 appears to excel in coding, scoring high in tests such as LiveCodeBench. According to Nature, unlike traditional 'reasoner' models, Kimi K2 is designed to execute complex multi-step actions using external tools autonomously, and its accessibility via API at low cost has spurred rapid adoption on platforms like Hugging Face. However, in an April 2025 ChinaTalk article, Lennart Heim noted that while Chinese AI models are likely to match US counterparts in performance, the latter retains a decisive edge in computing capacity, driven by more advanced AI chips and superior system integration at scale. Moreover, Gregory Allen, in a March 2025 report for the Center for Strategic and International Studies (CSIS) think tank, stated that DeepSeek trained its V3 model using 2.8 million graphics processing unit (GPU) hours on Nvidia H800 chips—export-compliant processors specifically designed to comply with the US's October 2022 chip controls. Allen noted that although DeepSeek's publications claimed exclusive use of H800s, reporting from SemiAnalysis and Chinese media—cited in the report—alleged that DeepSeek's R1 model may have been trained using banned Nvidia H100 chips. He reported that SemiAnalysis estimated DeepSeek's parent company, High-Flyer Capital, had acquired 50,000 Hopper-generation GPUs, including 10,000 H100s, 10,000 H800s, and 30,000 H20s. Allen further observed that Nvidia's A800 and H800 chips initially skirted US export controls until regulatory updates in October 2023 closed those loopholes. In addition to relying on US AI chips, China depends on US Electronic Design Automation (EDA) software for chip development. Reuters reported that the US has lifted export controls on EDA software, coinciding with China's relaxation of rare-earth export restrictions. Despite efforts toward indigenous advanced AI chip production—particularly Extreme Ultraviolet (EUV) lithography machines—the Hunan Printed Circuit Association noted this month that China remains at an early developmental stage, encountering significant issues with throughput, durability, and integration into existing ecosystems. These chip-related limitations may have sharp implications for the development of strategic Chinese platforms, notably the H-20 stealth bomber. According to the 2024 US Department of Defense (DoD) China Military Power Report (CMPR), the H-20 is a critical next-generation long-range bomber designed to bolster China's nuclear triad and extend military reach beyond the Second Island Chain. The report states that the H-20, based on a flying-wing design similar to the US B-2, is expected to surpass an 8,500-kilometer range and carry conventional and nuclear payloads, giving China its first true strategic bomber and global strike capability. It notes that the H-20 has yet to be revealed or flight-tested and may only enter service by the 2030s. In contrast to the H-20, General Thomas Bussiere, head of US Air Force Global Strike Command, stated in an interview this month with Air & Space Forces Magazine, that a second developmental B-21 bomber 'should fly shortly,' following the first unit's initial November 2023 flight. The magazine reports that production acceleration was enabled by Congressional approval of a $4.5 billion increase through a reconciliation bill, which Bussiere described as expected and based on over a year of analysis regarding capability, cost, and ramp-rate potential. Bussiere told the publication that this expansion reflects a growing recognition of the strategic value of long-range strike, particularly amid the challenge of sustaining aging Cold War-era bombers and an increasingly volatile global environment. While the official production goal remains 'more than 100' B-21 bombers, the article notes that Bussiere informed the Senate Armed Services Committee he supports assessing an increase to 145 aircraft, citing strategic shifts such as Russia's invasion of Ukraine and China's expanding strategic forces. Air & Space Forces Magazine adds that General Anthony Cotton, head of US Strategic Command, also advocates for raising the total to 145 aircraft. Highlighting the role of AI in accelerating B-21 development, Newsweek reported in December 2023 that AI optimized digital design and engineering processes, including simulation-based testing before physical construction. According to the report, AI-driven tools have enabled Northrop Grumman to maintain tight schedules, enhance sustainability, and optimize supply chains. It also notes that the B-21 incorporates open-architecture software, facilitating rapid upgrades and AI-driven mission flexibility, transforming it into a stealthy sensor and data fusion node beyond its bomber role. Despite these advantages, the US faces significant challenges scaling B-21 production. In a June 2025 report for the Heritage Foundation, Shawn Barnes and Robert Peters noted that the US relies solely on a single B-21 production facility in Palmdale, California, which limits output to about ten bombers per year—insufficient to reach the US Air Force's 100-aircraft goal before the late 2030s. They highlighted the high up-front development costs of the program, the fragility of the defense industrial base and the single-point-of-failure risk associated with relying on one site. Barnes and Peters argued that establishing a second production line is essential to scale capacity, reduce risk, and potentially support future sales to close allies, as was done with the F-35. AI-driven breakthroughs are clearly reshaping stealth aircraft development, but progress for both China and the US hinges critically on overcoming chip dependencies and scaling production capacities. The outcome of this high-stakes technological competition will shape the future balance of strategic airpower.
