DeepSeek-v4-flash-vision-exp
DeepSeek has launched its v4-flash-vision-exp model, bringing multimodal vision capabilities to its API via OpenAI and Anthropic-compatible endpoints. The detailed documentation outlines how to integrate images, from Base64 to external URLs, with specific limits and token usage for image processing. Hacker News discussion primarily revolves around the model's image resolution limitations, with many noting the 800x800 equivalent scaling might hinder detailed tasks like OCR.
The Lowdown
DeepSeek has released comprehensive API documentation for its new deepseek-v4-flash-vision-exp model, enabling developers to integrate image processing capabilities into their AI applications. This update allows the model to analyze images alongside text, opening doors for tasks like describing pictures, reading text from screenshots, and analyzing charts.
- The model supports common image formats: JPEG, PNG, GIF, and WebP.
- Images can be provided in three ways: Base64-encoded directly in the request, via a publicly accessible external URL, or by referencing a file uploaded through DeepSeek's Files API for larger or frequently reused images.
- Developers can control the image processing detail level, with
lowdetail downscaling images to 512x512 for faster, cheaper inference, whilehighororiginalretain the image's fidelity. - Images are converted into tokens for billing, with all images automatically resized to roughly the pixel count of an 800x800 image, resulting in an upper bound of 384 tokens per image.
- Strict limits are in place for request body size (48 MiB), single image size (up to 64 MiB via Files API), and images per request (up to 600).
- The vision capabilities are restricted to
usermessages and thedeepseek-v4-flash-vision-expmodel only. - The API also provides compatibility with Anthropic's
/messagesendpoint for image input.
This documentation serves as a critical guide for developers looking to leverage DeepSeek's new vision-enabled AI model, detailing integration methods, performance considerations, and usage constraints.
The Gossip
Resolution Ruckus
Many commenters expressed concern over the model's effective image resolution limit, noting that images are scaled down to roughly an 800x800 pixel equivalent. This limitation is seen as a significant drawback for tasks requiring fine detail, such as OCR on full-page documents or analyzing intricate charts, with one user comparing it unfavorably to 1995 Super VGA resolutions. The consensus is that while useful for some applications, higher resolutions are needed for broader utility.
Visionary Validation
The community largely welcomed the introduction of vision capabilities, with some users expressing relief and excitement. One commenter humorously noted that previous versions of the DeepSeek model sometimes 'invented' image analysis tools when they lacked actual vision, indicating that this official release is a much-needed and significant upgrade for the platform.
Cost & Capability Queries
Beyond resolution, discussions touched upon the economics and future of the model. One user calculated the token cost, estimating around 2,500 images per dollar based on the specified 400 tokens per image. There was also a common question regarding whether this vision model would be made open-weights, reflecting a general interest in accessibility and community contribution to AI models.