Hangzhou, China — DeepSeek, the Chinese AI lab backed by hedge fund High-Flyer, released DeepSeek V4 today, a trillion-parameter open model built on a highly efficient sparse architecture. The company, founded in July 2023 by Liang Wenfeng, continues its rapid cadence of releases that began with the DeepSeek-R1 model and its eponymous chatbot in January 2025.
DeepSeek V4 arrives as the latest salvo in an increasingly intense global race to build larger, more capable open-weight language models. The company’s choice to keep the model open — releasing its weights and architecture to the public — signals a bet that community-driven innovation and transparency can match or exceed the closed, proprietary strategies of rivals like OpenAI. DeepSeek-R1 had already demonstrated that a Chinese startup could produce responses comparable to GPT-4 and o1, and at a reported training cost significantly lower than those Western counterparts.
The sparse design at the heart of V4 is the technical headline. Rather than activating all trillion parameters for every query, the model dynamically selects only the relevant sub-networks needed for a given task. This approach, known as mixture-of-experts, has been pursued by labs including Google and Mistral, but DeepSeek appears to have scaled it to an unprecedented size. The result is a model that can, in principle, match the raw knowledge capacity of dense trillion-parameter models while requiring far less compute at inference time. For developers and researchers, that could mean running state-of-the-art AI on more modest hardware — or serving millions of users without bankrupting a data center.
DeepSeek’s rise has been one of the more surprising stories in AI. The company is owned and funded by High-Flyer, a quantitative hedge fund, giving it a financial backstop that most AI startups lack. Liang Wenfeng serves as CEO of both firms, and his background in high-frequency trading — an industry obsessed with latency and efficiency — may have informed the sparse architecture choices now making headlines. The lab operates out of Hangzhou, a tech hub already home to Alibaba and a dense ecosystem of AI talent.
What makes DeepSeek V4 particularly interesting is the timing. The model lands in a market where the cost of training frontier models has become a central anxiety for the industry. DeepSeek-R1’s relatively low training bill was widely discussed as a proof point that efficient engineering could undercut the massive capital expenditures of companies like OpenAI and Google. V4 extends that logic: if a trillion-parameter model can be built and run efficiently, the barriers to entry for the next generation of AI might be lower than many assumed.
The open-weight release also puts pressure on competitors who have been moving toward more closed, API-only models. Meta’s Llama series has been the most prominent open alternative, but DeepSeek V4 now offers a Chinese counterpart with comparable scale. For researchers in academia, startups, and countries without access to the latest Western models, V4 provides a powerful tool that can be inspected, fine-tuned, and deployed without licensing fees or API dependence.
Looking ahead, the arrival of DeepSeek V4 suggests that the frontier of open AI is not slowing down. If the sparse architecture delivers on its efficiency promises, it could accelerate a shift in how the entire field thinks about model scale — away from brute-force parameter counts and toward smarter, more selective computation. For Liang Wenfeng and his team, today’s release is a statement that the most ambitious AI work is no longer confined to Silicon Valley. The next phase of this story will be written in Hangzhou, and the model is now in the hands of the world.
