DeepSeek founder declares AGI as the goal as $10bn funding round advances

DeepSeek’s founder Liang Wenfeng told prospective investors that the Hangzhou lab will pursue artificial general intelligence as a primary goal. The lab plans to keep releasing open-source models rather than focus on short-term commercialization. The company is raising capital externally for the first time, with the reported valuation target around 70 billion yuan and a sought external amount of at least $300 million. Previously, the lab was funded entirely through High-Flyer Quant, which still effectively backs it. The move reflects the scale of training runs that exceed what a profitable hedge fund can self-finance. The company’s open-source V4-Pro and V4-Flash models and their optimization for Huawei Ascend and Cambricon silicon signal a strategy aimed at a domestic market with limited access to top US accelerators.

"DeepSeek founder Liang Wenfeng has told prospective investors in the company's first outside funding round that the lab intends to pursue artificial general intelligence as a primary goal, and will keep releasing open-source models rather than chase short-term commercialisation. Bloomberg reported the messaging on Friday, in coverage of an ongoing 70 billion yuan (roughly $10bn) raise."

"That figure represents the valuation DeepSeek is reportedly targeting, rather than the cheque size. The startup is seeking at least $300m in external capital in this round, according to. The Information has separately written that the longer-term ambition could reach $7bn or more as the lab moves toward recurring revenue. The current round is the first time Liang has accepted outside money."

"DeepSeek had until now been funded entirely out of High-Flyer Quant, the quantitative trading firm Liang founded and which still effectively backs the lab. Liang has previously framed the absence of outside investors as a deliberate insulation from product-roadmap pressure. The shift this round represents is partly a function of scale: training runs at the size DeepSeek now operates at have moved beyond what even a profitable hedge fund can self-finance."

"DeepSeek released its V4-Pro and V4-Flash models in April, a 1.6-trillion-parameter Mixture-of-Experts system and a smaller 284-billion-parameter variant, both released open-source under permissive licensing. The V4 family is optimised to run on Huawei Ascend and Cambricon silicon as well as Nvidia, which is a deliberate signal to a domestic Chinese market increasingly cut off from the highest-end US accelerators."

Read at TNW | Artificial-Intelligence