DeepSeek is accelerating its push into multimodal AI. On August 21, the company launched its experimental vision-language model, DeepSeek-V4-Flash-Vision-Exp, which adds image understanding capabilities on top of its existing text processing power, with API access now open to developers.

According to the company's published test results, the new model delivers performance on par with DeepSeek-V4-Flash across pure text tasks such as Agent reasoning, inference, and world knowledge. In vision-based Agent benchmark tests, however, it shows a significant improvement over the previous generation, drawing close to Opus-4.8's performance levels.

Alongside the model release, DeepSeek has rolled out a free Files API and published DeepSeek Harness 0.1.1 to provide integration support for the new offering. This update broadens the scope of DeepSeek's API suite, enabling developers to handle mixed image-text tasks within Agent frameworks.

Text Performance Steady, Visual Understanding Sharply Enhanced

Positioned as an experimental release, DeepSeek-V4-Flash-Vision-Exp focuses on preserving text capabilities while adding vision functionality.

DeepSeek reports that on pure text tasks involving Agent operations, reasoning, and world knowledge, the new model matches the performance of the V4-Flash production version. In Agent benchmarks requiring visual comprehension, it achieves a marked performance leap.

It is worth noting that DeepSeek conducted its testing using the DeepSeek Harness minimal-mode framework with parameters set to max tier, temperature=1.0, and top-p=0.95. Since V4-Flash disregards multimodal elements in the ApexBench and Agents' Last Exam tests, the comparison between the two generations on vision tasks is not entirely apples-to-apples.

Multimodal Design Tailored for Agent Use, API Pricing Unchanged

DeepSeek showcased the new model's versatility across various Agent scenarios, including generating a Tibet road trip presentation for high-net-worth clients based on user requirements, performing multi-round interactive edits on websites, and creating front-end mini demos with dynamic visual effects.

These demonstrations highlight the model's integrated understanding of both image and text data, as well as its potential in multi-turn interactive workflows.

On the API pricing front, V4-Flash-Vision-Exp retains the same fee structure as V4-Flash. Images are converted into token-based billing, with each image consuming up to 384 tokens. The API supports three calling formats—Chat Completions, Messages, and Responses—and images can be submitted via base64 encoding, external URLs, or the Files API.

The newly introduced Files API is currently free of charge. Developers can pre-upload images and then reference them by file_id in subsequent requests to avoid redundant uploads. DeepSeek has also released DeepSeek Harness 0.1.1, offering out-of-the-box Agent integration support for the new model.