- DeepSeek has launched its new artificial intelligence model, DeepSeek V4, claiming world-beating capabilities and optimisation for domestically produced Chinese chips.
- The model, available in V4-Pro and V4-Flash versions, features an ultra-long context of one million words, leading in agent capabilities, world knowledge, and reasoning among open-source models.
- DeepSeek V4-Pro is said to significantly outperform other open-source models and is only slightly behind Google's Gemini-Pro-3.1 in world knowledge benchmarks.
- The release follows DeepSeek's R1 model, which last year triggered a trillion-dollar stock market sell-off in the Western tech industry due to its cost-effectiveness and performance rivalling US giants.
- This new version dramatically improves computational efficiency, processing up to one million tokens, enabling multi-document reasoning and coming amidst US semiconductor export restrictions on China.
IN FULL