DeepSeek has presented an experimental artificial intelligence model capable of processing visual queries. According to the company, its performance is approaching that of the leading model of its American competitor – Anthropic. The new product is an experimental version of the flagship text model DeepSeek V4 Flash, with the addition of multimodal capabilities: it is capable of analyzing and processing visual queries, including images and screenshots. The model is called DeepSeek‑V4‑Flash‑Vision‑Exp and is already available on the company’s API platform.

The experimental multimodal model is on par with DeepSeek-V4-Flash in terms of text processing, including agentic tasks, logical reasoning, and general world knowledge. In tests of multimodal agentic capabilities, the new model makes a significant leap forward compared to the base version, approaching Anthropic’s Opus 4.8 in performance. Images are charged at up to 184 tokens per image at V4-Flash prices, and the model supports mixed text and image input via base64, external URLs, or the Files API.

Developers from different countries are actively competing in the field of artificial intelligence, offering models with similar performance at a lower price. The emergence of the flagship DeepSeek model in early 2026 changed the perception of the capabilities of affordable open‑weight models.