DeepSeek Releases Experimental V4-Flash-Vision Model
DeepSeek has released V4-Flash-Vision-Exp, an experimental vision model built on V4-Flash that can parse diagrams, recognize text, and call tools. By DeepSeek’s own measurements it performs at the level of Claude Opus 4.8.
Images can be passed as Base64, via public URL, or through the new Files API (upload once, reference by ID). Formats are detected by actual file content rather than extension, one image costs at most 384 tokens, and prompts accept up to 600 files. Images are normalized to 800x800 pixels, with a detail parameter to force 512x512; the max side length is 8192 pixels (dropping to 4096 with 15+ images).
Related: DeepSeek Launches V4-Flash API with Agentic Upgrades, DeepSeek Upgrades V4-Flash (0731)