Deepseek has released an experimental Flash Vision model that demonstrates strong performance on agent-related tasks. The model, named V4-Flash-Vision-Exp, adds image understanding to the company's existing text capabilities. On Deepseek's internal benchmarks, the vision model nearly matches Opus 4.8 in agent tasks, according to the company. It extends the text-only Deepseek-V4-Flash model with image processing while maintaining the base model's performance in reasoning and world knowledge. The vision variant scores close to Opus 4.8 on Deepseek's internal multimodal agent benchmarks, the company said.
The model is designed for agent-based applications and works with different agent frameworks, combining visual understanding with tool use. In practice, it can describe images, extract text from screenshots, and analyze diagrams. It supports JPEG, PNG, GIF, and WebP formats, determining the format from actual file content rather than filename or declared MIME type, as stated in the API documentation. The model also works with OpenAI's Chat Completions and Responses APIs and Anthropic's Messages endpoint, according to Deepseek.
Deepseek has also released version 0.1.1 of its Harness framework, which supports the new model out of the box. The model allows developers to send images via Base64 encoding, public URLs up to 32 MiB, or the new free Files API with a 64 MiB size limit. An optional 'detail' field downscales images to 512 x 51,2 pixels, saving tokens when fine detail isn't needed. The model automatically normalizes images to around 800 x 800 pixels based on aspect ratio before processing. Each image costs at most 384 tokens, regardless of original resolution. A single request can include up to 600 images, with a maximum edge length of 8,192 pixels per side, dropping to 4,096 pixels once a request contains 15 or more images. Images can only go in user messages.
Source: thedecoder
