New model from DeepSeek to heat up the AI race

China-based artificial intelligence company DeepSeek has introduced its new experimental model capable of understanding visual inputs. The Hangzhou-based company announced DeepSeek-V4-Flash-Vision-Exp, which is positioned as a version of the text-only DeepSeek-V4-Flash model with added multimodal capabilities.

The new model can analyze visual inputs such as images and screenshots in addition to text prompts. According to the company's statement, the model can perform specific tasks using the information it extracts from these images.

In a post on X, DeepSeek stated that the experimental model maintains the same level as DeepSeek-V4-Flash in terms of text capabilities. The company reported that the model matches its current flagship model in areas such as agent features, reasoning, and world knowledge.

According to the company's claims, DeepSeek-V4-Flash-Vision-Exp showed performance "close" to Anthropic's advanced Opus 4.8 model in tests measuring multimodal agent capabilities. Agent capabilities refer to an AI model's ability to follow and complete assigned tasks without constant human guidance.

DeepSeek had previously worked on visual and multimodal models such as the DeepSeek-VL family. The newly announced model stands out for adding the capacity to interpret images and execute tasks based on these inputs to the company's most advanced model series.

DeepSeek-V4-Flash-Vision-Exp is accessible through the company's API service. The announcement comes at a time when competition between Chinese AI companies and US-based developers is intensifying.

Chinese developers have recently been drawing attention with models that offer powerful performance at a lower cost. The flagship model that DeepSeek released at the beginning of the year also changed expectations regarding the capacity of open-weight and low-cost AI systems.

News Source: 12punto