DeepSeek expands AI capabilities with new vision model for image processing
The deepseek-v4-flash-vision-exp model allows developers to process images alongside text, supporting multiple formats and integrating with existing OpenAI and Anthropic API frameworks.
DeepSeek has introduced the deepseek-v4-flash-vision-exp model, marking a significant shift from its previous text-only offerings to multimodal capabilities. The new model accepts images alongside text, enabling users to describe pictures, extract text from screenshots, and analyse charts. This release is documented in the company’s API guides, which were highlighted on the Hacker News community platform on 21 August 2026.
The model supports JPEG, PNG, GIF, and WebP file formats, detecting the format from the actual file content rather than the file name or declared MIME type. Developers can provide images to the model through three distinct methods: base64 data URLs, public http(s) links, or the Files API. All methods utilise the standard OpenAI-compatible Chat Completions format, where content is structured as an array of blocks rather than a plain string.
Technical specifications indicate that inline images, whether sent via base64 or file data, count toward a 48 MiB request body limit. For public links, the URL must be at most 8192 characters, the image file may be up to 32 MiB, and the download must complete within 60 seconds. Images referenced via the Files API may be larger, up to 64 MiB, and are not subject to the 32 MiB per-image check, making this option suitable for reusing images across multiple requests or handling larger files.
To manage computational costs, every image is automatically resized before inference. This process caps token usage at 384 tokens per image, regardless of the original dimensions. Consequently, a 2000 by 2000 pixel image and a 5000 by 5000 pixel image consume the same number of tokens after resizing. These image tokens are billed alongside text tokens, and when a request contains multiple images, each is counted independently under the same rule.
The model is designed for seamless integration into existing developer workflows. It is accessible via OpenAI-compatible Chat Completions and the Responses API, as well as an Anthropic-compatible /messages endpoint. This compatibility allows developers already using similar frameworks to adopt the new vision capabilities without significant architectural changes. The "exp" designation in the model name suggests it may be an experimental or preview release, and specific pricing details for vision tokens are not yet provided beyond the general billing structure.
