DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, the first model in its V4 series to support image input.
This is a multimodal model with 305 billion parameters. It incorporates an image processing module based on the DeepSeek-V4-Flash architecture and adds image recognition capabilities through post-training.
In tests including image recognition tasks, the model recorded scores equivalent to those of Claude Opus 4.8. Regarding text processing performance, it maintains performance equal to or greater than that of DeepSeek-V4-Flash-0731.
DeepSeek-V4-Flash-Vision-Exp is available for download from Hugging Face and ModelScope under the MIT License.
It is also available via API. The supported image formats are JPEG, PNG, GIF, and WebP.
The maximum resolution for input images is 8,192 pixels on the longest side; however, if 15 or more images are input, the longest side is limited to 4,096 pixels. Since large images are resized to an equivalent of 800x800 pixels while maintaining the aspect ratio, the number of tokens per image is approximately 384 tokens maximum.
Source: DeepSeekが画像認識に対応した「DeepSeek-V4-Flash-Vision-Exp」をオープンモデルとして公開、画像認識を含むタスクでClaude Opus 4.8と同等の性能 (GIGAZINE, 2026-09-02)