DeepSeek V4: Cheaper Frontier Models With a Million-Token Window
DeepSeek’s new V4 model signals how aggressively China is competing in the AI race. The Hangzhou startup has launched DeepSeek V4-Pro and DeepSeek V4-Flash, positioning them as cost‑efficient alternatives to US models. V4-Pro is the larger system for demanding reasoning tasks, while V4-Flash uses fewer parameters and is tuned for speed and lower operating costs. Both support an ultra‑long one‑million token context window, allowing the model to handle entire codebases, lengthy research documents or complex project histories in a single prompt. DeepSeek claims V4-Pro significantly outperforms other open-source peers on world-knowledge benchmarks and trails only Google’s Gemini‑3.1‑Pro among top closed models. Crucially, V4 is designed for “drastically reduced compute and memory costs,” directly addressing the bottleneck created by today’s AI compute crunch. For developers, that makes the DeepSeek V4 model one of the first truly cheap frontier models with mainstream‑ready long‑context capabilities.
DeepSeek as China’s Efficient Challenger to US AI Leaders
DeepSeek has rapidly become one of the most closely watched AI startups in China by attacking the cost side of frontier models. Its earlier R1 reasoning model surprised observers by matching US rivals at significantly lower cost, upending assumptions about permanent US dominance. With DeepSeek V4, the company is doubling down on compute efficiency, marketing “world‑leading long context” at reduced hardware and memory requirements. V4-Pro reportedly uses 1.6 trillion parameters, while V4-Flash runs on 284 billion, and both are optimised for popular AI agents such as Claude Code, OpenClaw and OpenCode. A preview version is openly downloadable on platforms like Hugging Face, extending its reach beyond DeepSeek’s own chatbot. The ability to run V4 on Huawei’s Ascend hardware, despite US sanctions on Huawei, underlines a parallel Chinese hardware‑software stack. Together, these moves position DeepSeek as a flagship of the China AI race, proving that global‑class models can be delivered more cheaply and more openly.
China Blocks Meta’s Manus Acquisition on National Security Grounds
While DeepSeek rises at home, Beijing is also tightening its grip on outbound AI assets. China’s National Development and Reform Commission has ordered the unwinding of Meta’s planned acquisition of Manus, a high‑profile AI startup founded in China and based in Singapore. The regulator said it would prohibit foreign investment in Manus on national security grounds and required all parties to withdraw from the transaction. Manus gained global attention after unveiling what it called the “world’s first” fully autonomous AI agent, capable of independently executing tasks from holiday bookings to game development. Co‑founder Yichao Ji framed the platform as “the next paradigm of human‑machine collaboration” and a potential glimpse of AGI, quickly attracting a waiting list of over two million users. By blocking the Meta Manus acquisition, Beijing is signalling that strategic AI talent and intellectual property will not be easily absorbed into US tech giants, especially amid rising AI geopolitical tensions.
A Coherent Strategy: Back Local Champions, Limit Strategic US Deals
Viewed together, DeepSeek V4’s launch and the blocked Meta Manus acquisition reveal a broader Chinese strategy in the AI race. Domestically, regulators and industry are creating room for local champions like DeepSeek to thrive with cheap frontier models and tight integration with Chinese hardware ecosystems. Internationally, authorities are constraining foreign access to high‑impact AI assets, especially those involving autonomous agents or key research teams. Beijing has reportedly described the Manus deal as an attempt to hollow out its tech base and is tightening oversight of AI firms that seek to leave the country. This dual track—promote efficient, world‑class models at home while limiting strategic acquisitions by US platforms—reshapes competitive dynamics. It pushes US firms to either build similar capabilities from scratch or partner under stricter Chinese terms, deepening AI geopolitical tensions and reinforcing national boundaries around data, compute and talent.
Implications for Asia’s Developers, Enterprises and Regulators
For developers and enterprises across Asia, these shifts present both opportunity and complexity. DeepSeek V4’s lower compute and memory requirements could significantly reduce the cost of long‑context applications, from legal and financial analytics to large‑scale code refactoring, especially where access to top US models is constrained or expensive. As Chinese and US ecosystems diverge, organisations may face a strategic choice between DeepSeek‑style models and US offerings, balancing price, performance and data‑sovereignty concerns. Open access to the DeepSeek V4 model on platforms like Hugging Face may appeal to startups that want local deployment and tighter data control. At the same time, regulators in the region must navigate competing standards and rising AI geopolitical tensions, deciding how to treat cross‑border deals like the Meta Manus acquisition. In the wider race to build powerful world models and autonomous agents, Asia’s policy choices will help determine whose systems developers can most easily build on—and at what cost.