Chinese AI startup DeepSeek has released a preview version of its new artificial intelligence model, DeepSeek V4, claiming it outperforms all open-source competitors. The company stated that the model features an ultra-long context of one million words, achieving leadership in agent capabilities, world knowledge, and reasoning performance across domestic and open-source fields.
DeepSeek V4 is available in two versions: V4-Pro and V4-Flash, with the latter described as a more efficient and economical choice. The company noted that V4-Pro significantly leads other open-source models in world knowledge benchmarks, only slightly outperformed by Google's closed-source Gemini-Pro-3.1. It also includes a 'maximum reasoning effort mode' that DeepSeek claims advances open-source knowledge capabilities.
The release follows DeepSeek's R1 model, which triggered a trillion-dollar stock market sell-off last year by rivalling systems like OpenAI's ChatGPT at a fraction of the cost. The R1 launch caused Nvidia to lose over $500bn in market value and impacted other tech giants. DeepSeek's models are open-source and free to use, marking a significant challenge to US AI dominance.
DeepSeek V4 is reportedly better optimised for domestically produced chips, amid growing US semiconductor export restrictions. The company did not disclose the chip system used for training but said its software works with both Nvidia and Huawei chips. The new model can process up to 384,000 tokens in output and understand context of up to 1 million tokens, a leap from V3's 128,000-token capacity.
This upgrade enables multi-document reasoning, allowing the AI to understand entire books and code databases. DeepSeek claims V4-Pro outperforms Google's Gemini-3.1-Pro in long-context understanding but remains behind Anthropic's Claude Opus 4.6. The company aims to further enhance the model's intelligence and usability across various tasks.
