According to CoinMeta, DeepSeek has officially launched a new multimodal model, deepseek-v4-flash-vision-exp. This model is on par with deepseek-v4-flash in terms of text capabilities and covers areas such as agent, reasoning, and world knowledge. In multimodal agent benchmark tests, v4-flash-vision-exp has made significant progress compared to v4-flash, bringing the performance of multimodal agents closer to that of opus-4.8.