Chinese AI startup DeepSeek on Friday released a preview of its next-generation model, DeepSeek-V4. The company said the model supports an ultra-long context of one million words.
DeepSeek-V4 is offered in two variants: V4-Pro and a lower-cost V4-Flash. V4-Pro has 1.6 trillion parameters, while V4-Flash has 284 billion parameters. In a company statement, DeepSeek said V4-Pro “significantly leads other open-source models” on world knowledge benchmarks and is only narrowly behind a leading closed-source model, Google’s Gemini-Pro-3.1.
The firm said the preview release is intended to gather real-world feedback before the model is finalized.
US-China AI race
DeepSeek garnered attention in January last year after unveiling a generative AI chatbot whose capabilities rivaled US products like ChatGPT, which DeepSeek said had required far less compute and funding to develop. The company has also faced controversy: its chatbot has been reported to avoid questions on politically sensitive subjects such as the 1989 Tiananmen crackdown, raising censorship concerns.
The Hangzhou-based startup has been accused by the United States and some competitors of improper or illegal conduct. The White House recently alleged that Chinese entities were conducting “industrial-scale distillation campaigns to steal American AI.” Beijing denied the claims, calling them baseless and saying China attaches great importance to protecting intellectual property rights.
Edited by: Sean Sinico