DeepSeek unveiled an experimental addition to its V4 Flash family that extends the model's capabilities beyond text to include visual prompts, the Hangzhou-based developer said. The release, named DeepSeek-V4-Flash-Vision-Exp, is being distributed through the DeepSeek API Platform.
According to the company, the Vision-Exp variant preserves the text competencies of DeepSeek-V4-Flash - including agent functionality, reasoning, and general world knowledge - while adding the ability to process images and screenshots. DeepSeek reported that on multimodal agent benchmarks the new model shows a substantial improvement over the text-only V4-Flash.
DeepSeek said that the multimodal agent performance of V4-Flash-Vision-Exp approaches that of Opus-4.8. The comparison refers specifically to agentic multimodal capabilities, defined as a model's capacity to act with reduced need for constant oversight or prompting.
On input handling, the company stated that images are tokenized for billing purposes at up to 384 tokens per image and charged at V4-Flash pricing. The model supports mixed text-and-image input through base64 encoding, external URLs, or via the Files API.
The announcement framed the release in the broader context of competition between Chinese AI firms and U.S. developers, noting that models from Chinese companies frequently deliver comparable performance at a lower cost. DeepSeek's earlier launch of its flagship model this year, the company said, recalibrated expectations for what lower-cost, open-weight models can accomplish.
DeepSeek positioned the Vision-Exp build as experimental. The company emphasized parity with its existing text model on textual tasks and pointed to measurable gains on multimodal benchmarks versus the prior V4-Flash baseline. Beyond those benchmark claims and the billing details, DeepSeek provided no additional technical metrics in its announcement.
Technical availability: DeepSeek-V4-Flash-Vision-Exp is available now on the DeepSeek API Platform and accepts mixed media inputs as described above.