Chinese artificial intelligence (AI) startup DeepSeek has unveiled a new model that can process up to 1 million tokens at once.
DeepSeek on the 24th (local time) said on social media (SNS) that it "will officially release a preview of a completely new series of models, 'DeepSeek-V4,'" unveiling V4 Flash and V4 Pro.
V4, built on a greatly expanded context compared with existing large language models, can process long documents and codebases of up to 1 million tokens at once. One million tokens equal about 700,000 to 800,000 English words, equivalent to seven to 10 full-length novels.
The performance-focused V4 Pro aimed for capability on par with closed-source models. DeepSeek said V4 Pro is among the best open-source models in agent execution and that its overall knowledge and reasoning abilities are similar to those of top-tier global closed models. However, its knowledge performance is somewhat lower than Google's "Gemini Pro 3.1."
The lightweight V4 Flash focuses on price competitiveness. While its overall knowledge level is lower than V4 Pro, its reasoning ability is at a comparable level, according to DeepSeek.
The performance-focused V4 Pro is priced at $1.74 for input and $3.48 for output per 1 million tokens (about 2,600 won and about 5,100 won, respectively). V4 Flash is lower at $0.14 for input and $0.28 for output (about 200 won and about 400 won). This is significantly lower than major competing models.
According to the industry, OpenAI's GPT series, Google Gemini, and Anthropic Claude vary by model, but typically cost about $10 to $30 for input and $30 to $60 for output per 1 million tokens (about 15,000 to 45,000 won and about 45,000 to 89,000 won). In other words, DeepSeek V4 Flash has reduced the expense to as little as one-tenth and up to one-fiftieth compared with competing models.
Industry watchers say DeepSeek is seeking to expand market share by leveraging price competitiveness now that it has narrowed the performance gap to a certain level. If infrastructure based on Huawei's Ascend 950 chips is built, prices could fall further.
However, Bloomberg said DeepSeek may have performed "distillation," training the new model using responses from other AI models such as OpenAI and Anthropic as data, and there are also suggestions it used Nvidia's advanced chips, which are banned for export to China.
DeepSeek reshaped the AI industry in January last year by releasing the low-cost, high-performance model R1. At the time, amid concerns that DeepSeek's popularity could burst the AI bubble, U.S. tech stocks fell sharply.