DeepSeek has introduced an experimental multimodal AI model that can understand visuals and add this capability to its text-only V4 Flash series. Through this, they aim to create a strong system and compete with Western models.
A Hangzhou Chinese startup has put out DeepSeek-V4-Flash-Vision-Exp on August 21, 2026, and now the public can access this model through the DeepSeek API platform. This new visual model is built on the fast-working V4 Flash architecture directly by the company. It can also process text and act by using the company’s own text-based model.
Beside text-only DeepSeek-V4-Flash performance is good for tasks like creating a chain of actions, reasoning, and using common knowledge of the world. Based on a DeepSeek announcement about their tests with multimodal agent benchmarks (how well a model reads a picture or chart and carries out a series of tasks without much human input), they have gone through a big transformation so fast that their results are almost equal to Anthropic Opus 4.8. On the 11 tasks they have tested their new model, the results have outperformed Opus 4.8 on three tasks still they lagged behind on the rest.
DeepSeek, in the past, have put out more than one multimodal systems under DeepSeek-VL umbrella. This however is the one where vision has been seamlessly added to a DeepSeek model. DeepSeek has also upgraded the open-source DeepSeek Harness tool (version 0.1.1) to include, without effort, the new model, and the goal here is, to make it possible for agent setups to merge the understanding of a visual with the execution of a task in the workflow.
DeepSeek has continued to launch new products swiftly after DeepSeek API release. When the R1 model became publicly available, which is a very price-competitive feature of DeepSeek models, DeepSeek drew international attention in the early days of 2025. With the V3 and V4 models, DeepSeek continued to emphasize open source distributions, affordability and, most importantly, robustness in reasoning and agency. This vision-only model of DeepSeek came at a time when the company is not only expanding the capabilities of its offerings but also engaging in a global competition with major American firms like Anthropic, OpenAI and Google against a backdrop of an ever-growing and highly competitive international development of AI.
Anyone interested can try DeepSeek-V4-Flash-Vision-Exp by contacting it through DeepSeek API and using model name deepseek-v4-flash-vision-exp. This experimental model may be subject to revisions as further tests of multimodal agent tasks are performed by users in various real-world contexts.