DeepSeek-v4-flash-vision-exp
Points and comments are a snapshot, not live.
DeepSeek releases the v4-flash-vision-exp model for image analysis via text prompts.
The deepseek-v4-flash-vision-exp model accepts images in JPEG, PNG, GIF, and WebP formats. Images can be provided as base64 data URLs, external HTTP(S) links (up to 8192 characters, 32 MiB), or via the Files API (up to 64 MiB). The model resizes images to roughly 800×800 pixels, capping token usage at 384 tokens per image. A detail parameter (low, high, original, auto) controls processing. Limits include 48 MiB request body, 600 images per request, and 8192 px per side (drops to 4096 px with 15+ images). Supports OpenAI, Anthropic, and Responses API formats.
What commenters are saying
Commenters welcome the update as DeepSeek v4 Flash 0731 often incorrectly assumed it had vision and tried to analyze images. The 800×800 resize limit is widely criticized; many argue it kills use cases like OCR and detailed diagrams. Suggestions include splitting images into overlapping crops, using zoom tools, or smaller models for layout detection (e.g., PP-DocLayoutV3). One user estimates 2,500 images per dollar. Some ask whether the model will be open-weights; others note DeepSeek previously stated multimodal support was planned, not excluded.