During the domestic Spring Festival holiday a year ago, OpenAI released its large video generation model Sora on February 15, 2024, local time. The seamless camera movements and almost lifelike presentation in several Sora-generated videos left the domestic large model industry, still in its imitation and following stage, in shock and pessimism. The notion of "surrender" was rampant, with investors and major companies advising entrepreneurs to abandon illusions and focus on applications, declaring that starting a business in large models was a "dead end."
Who would have thought that just a year later, during this Spring Festival, the discussion would be about a domestic large model called DeepSeek. Beyond the tech circles and viral discussions, its applications have begun to penetrate households, with more ordinary people using DeepSeek to customize diet plans, edit holiday greetings, write acrostic poems, and even tell fortunes.
So far, DeepSeek has launched three generations of models. In May last year, DeepSeek, under the umbrella of Phantom Square Quantitative, released DeepSeek-V2, claiming capabilities comparable to GPT-4 but at only about 1% of GPT-4's price, sparking a year-long price war in the domestic large model market. By December, DeepSeek released the new large model DeepSeek-V3, reducing training costs to a few million dollars, earning the title of "price butcher." The latest release, DeepSeek-R1, directly competes with OpenAI's o1. The launch of "deep thinking" and "network search" features propelled DeepSeek to the top of the free charts in both China and the US.