A research report from CICC indicates that DeepSeek recently launched its new-generation open-source V4 model series and published a technical report. As a representative domestic open-source model developer, DeepSeek's technological innovations are a key driver of progress in the model industry. The most significant architectural innovation in the V4 series is the application of a hybrid attention mechanism to reduce FLOPs per token and KV Cache, thereby improving model inference efficiency. The report suggests that DeepSeek and other domestic models are exploring methodologies for model advancement and engineering optimization through open-source approaches, collectively fostering industry growth and accelerating the arrival of the AGI era. CICC notes that the release of the V4 model by DeepSeek overcomes efficiency bottlenecks in ultra-long context processing. The V4 preview introduces two MoE model versions: V4-Pro, which offers higher intelligence and pricing for complex tasks, and V4-Flash, which emphasizes cost-effectiveness. Both versions support an ultra-long context of 1 million tokens, a core highlight of the V4 series. The report believes that V4's breakthrough in long-context capabilities will accelerate progress in Agentic AI, enabling more efficient and accurate completion of complex, long-range tasks. CICC further states that V4's optimizations alleviate HBM pressure, efficiently unlock storage efficiency across components, and highlight the trend of SSD integration into core model inference. The firm expresses optimism about the advancement trajectory of domestic open-source large models, noting that efficiency improvements will accelerate the release of downstream Agentic AI demand. It maintains a positive outlook on model developers KNOWLEDGE ATLAS (02513) and MINIMAX-WP (00100).
Comments