Chinese artificial intelligence company DeepSeek has previewed its new large language model, V4, as the global AI race accelerates.
DeepSeek released two versions: V4-Pro, a larger model for demanding tasks, and V4-Flash, a smaller, faster, and more cost-effective option.
DeepSeek claims V4-Pro leads open-source models in world knowledge benchmarks, closely trailing Google's Gemini-3.1-Pro. The model boasts a one-million token context length, enabling it to process extensive prompts like lengthy documents or code.
This capability allows the AI to consider significantly more text at once, leading to more comprehensive outputs. DeepSeek highlights "world-leading long context with drastically reduced compute and memory costs."
The new AI model integrates with popular AI agents, facilitating software development tasks. DeepSeek maintains its reputation for an open approach, making V4 freely available for testing and adaptation on platforms like Hugging Face.
The company previously gained attention for its open-source V3 model, noted for strong performance at a lower cost and with less powerful hardware than many US competitors.
The release of V4 coincides with OpenAI's launch of GPT-5.5 and follows accusations from the White House concerning China's alleged IP theft in AI. Several countries, including Italy, the US, and South Korea, had previously restricted DeepSeek usage over national security concerns, while Germany cited illegal data transfer issues.
This development occurs amidst warnings from major AI firms like OpenAI, Anthropic, and Google about Chinese AI companies employing "distillation" methods to replicate model capabilities.