DeepSeek Launches V4-Flash-Vision-Exp Multimodal Model and Files API
DeepSeek officially announced that the experimental multimodal model DeepSeek-V4-Flash-Vision-Exp is now available on the API platform. The model maintains parity with V4-Flash in text understanding, agent scheduling, and world knowledge, while achieving significant breakthroughs in multimodal agent benchmark tests, with overall performance approaching Opus-4.8. This update simultaneously opens the image-text mixed input interface, clarifies image token billing standards, and launches the accompanying DeepSeek Harness 0.1.1 version. Additionally, the platform has officially launched the free Files API, supporting one-time upload and multiple reuse of images, effectively reducing request bandwidth consumption.
DeepSeek recently announced that the experimental multimodal model DeepSeek-V4-Flash-Vision-Exp has been officially deployed to the DeepSeek API platform. Positioned as a multimodal extended version, it retains the original efficient inference architecture while significantly enhancing visual understanding capabilities. According to official test data, it maintains exactly the same level as DeepSeek-V4-Flash in core dimensions such as text generation, agent task scheduling, and world knowledge Q&A.
In specialized multimodal agent benchmark tests, V4-Flash-Vision-Exp demonstrates a significant generational leap. Compared to the previous generation pure text model V4-Flash, the new version greatly improves efficiency in handling complex visual instructions and tool call coordination, with overall performance metrics approaching Opus-4.8. This breakthrough provides more reliable underlying support for the implementation of cross-modal workflows.
At the API invocation level, developers can directly enable the model by setting the model parameter. The interface fully supports Chat Completions and Messages & Responses protocols, allowing users to submit text and image data together in a single request. Image input supports Base64 encoding, external URL links, and the newly added Files API reference method. Regarding billing rules, each image is calculated at a maximum of 384 tokens, with rates consistent with the V4-Flash standard.
To facilitate ecosystem adaptation for the new model, DeepSeek simultaneously released Harness 0.1.1. This toolchain natively supports V4-Flash-Vision-Exp out of the box, allowing developers to quickly complete model integration and performance tuning without additional configuration. Mainstream agent frameworks have already achieved seamless integration, further expanding scenarios combining visual perception and automated operations.
Meanwhile, the DeepSeek API platform has officially launched the Files API feature module. This service is freely available to all users, allowing developers to upload frequently used image files to cloud storage in one go. Subsequent requests only need to pass the corresponding file_id for direct invocation, and the system will automatically parse and link historical files. This not only avoids network delays caused by repeatedly transmitting the same assets but also significantly reduces the bandwidth usage and invocation costs of API requests.
KEY HIGHLIGHTS
Key highlights
Experimental Multimodal Model DeepSeek-V4-Flash-Vision-Exp Officially Integrated into API Platform
Text Capabilities Match V4-Flash, Multimodal Agent Performance Approaches Opus-4.8
Supports Chat Completions and Messages & Responses Interfaces, Compatible with Image-Text Mixed Input
Images Billed at a Maximum of 384 Tokens, Priced Consistently with V4-Flash
DeepSeek Harness 0.1.1 Version Provides Out-of-the-Box Support for New Models
Files API Launched Simultaneously, Supporting One-Time Upload and Repeated Calls via file_id