Chinese AI startup DeepSeek has released a preview 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 tokens, achieving leadership in agent capabilities, world knowledge, and reasoning performance.
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 and is only slightly outperformed by Google's Gemini-Pro-3.1, a top-tier closed-source model.
The release comes amid ongoing US semiconductor export restrictions, which have forced China to rely on domestic chip manufacturers. DeepSeek did not disclose the chip system used to train V4 but said its software components are compatible with both Nvidia and Huawei chips.
DeepSeek V4-Pro includes a 'maximum reasoning effort mode' and can process up to 384,000 output tokens. The model's one-million-token context length enables multi-document reasoning, allowing it to understand entire books and full code databases. The company claims this represents a dramatic leap in computational efficiency compared to the previous V3 model, which handled 128,000 tokens.
DeepSeek previously caused a trillion-dollar stock market sell-off with its R1 model, which rivalled systems like OpenAI's ChatGPT at a fraction of the cost. The V4 release aims to further solidify China's position in the global AI race.
