Chinese AI startup DeepSeek on Friday released a preview of its next-generation model, DeepSeek-V4, saying the system can handle an ultra-long context of up to one million words.
DeepSeek-V4 will be available in two configurations: V4-Pro and a lower-cost V4-Flash. The company reported that V4-Pro contains about 1.6 trillion parameters, while V4-Flash contains roughly 284 billion parameters. In its statement, DeepSeek said V4-Pro outperforms other open-source models on world-knowledge benchmarks and trails a leading closed-source model, Google’s Gemini‑Pro‑3.1, by a narrow margin.
DeepSeek framed the preview as a way to collect real-world feedback and usage data before finalizing the model for a broader release.
Background and controversies
The Hangzhou-based startup drew international attention in January last year after unveiling a generative AI chatbot whose performance DeepSeek said matched U.S. offerings such as ChatGPT, while claiming it required less compute and funding to develop. The company has also faced scrutiny for its chatbot’s handling of politically sensitive topics; independent reports said the system avoids questions about events such as the 1989 Tiananmen crackdown, prompting concerns about censorship.
DeepSeek has been accused by some U.S. officials and industry rivals of improper or illegal behavior. The White House has alleged that Chinese entities are conducting “industrial-scale distillation campaigns to steal American AI.” Beijing rejected those allegations, calling them baseless and saying China takes intellectual property protection seriously.
The company did not provide a timeline for a full release or pricing details beyond the two announced variants.
Edited by: Sean Sinico