Table of Contents
Chinese AI company DeepSeek has expanded its DeepSeek V4-Flash model family with an experimental multimodal version that can process images alongside text.
Called DeepSeek-V4-Flash-Vision-Exp, the model is now available through the DeepSeek API. DeepSeek describes it as an experimental vision-understanding model that retains the text capabilities of V4-Flash while adding visual input.
The release gives developers access to visual capabilities through the V4-Flash line without presenting the model as a new flagship system.
What Is DeepSeek V4-Flash Vision Exp?
The new model builds on DeepSeek V4-Flash, retaining its capabilities in areas including agents, reasoning and general knowledge, according to DeepSeek.
The key addition is image understanding. Developers can provide images alongside text and use the model for tasks such as describing pictures, reading text from screenshots and analyzing charts. DeepSeek’s documentation lists JPEG, PNG, GIF and WebP among the supported image formats.
The standard V4-Flash and the new vision variant should therefore be distinguished: DeepSeek-V4-Flash-Vision-Exp is the experimental model that adds visual input.
DeepSeek V4-Flash Vision Exp vs Claude Opus 4.8
DeepSeek has highlighted the new model’s performance on multimodal agent benchmarks, saying it brings performance close to Anthropic’s Claude Opus 4.8.
However, the comparison needs to be viewed carefully because the results are based on DeepSeek’s own evaluations rather than independent testing.
| Benchmark | DeepSeek V4-Flash Vision Exp | Claude Opus 4.8 |
| ApexBench | 36.5 | 39.4 |
| Agents’ Last Exam | 27.3 | 25.7 |
| ZeroBench | 35.0 | 34.0 |
| Terminal Bench 2.1 | 83.9 | 85.0 |
| NL2Repo | 57.7 | 69.7 |
The figures show that the models trade results across different evaluations. DeepSeek V4-Flash Vision Exp is ahead on some tests but behind on others. For example, it trails Claude Opus 4.8 on ApexBench and NL2Repo while scoring higher on Agents’ Last Exam and ZeroBench.
DeepSeek’s published benchmark methodology uses its Harness Minimal Mode, and the results have not been independently verified. Therefore, the figures should not be interpreted as evidence that the experimental model universally matches or surpasses Claude Opus 4.8.
API Availability and Image Processing
Developers can access the model through the DeepSeek API using the identifier:
deepseek-v4-flash-vision-exp
DeepSeek says images are converted into tokens based on their dimensions and billed as input tokens. The company has capped image processing at 384 tokens per image.
The model uses the existing V4-Flash pricing structure, according to DeepSeek’s announcement, allowing developers to access the new visual capability within the same pricing framework.
DeepSeek also released Harness 0.1.1 with support for the experimental model.
New Files API for Developers
Alongside DeepSeek V4-Flash Vision Exp, DeepSeek introduced a Files API that allows developers to upload an image and reference it again through a file ID.
The Files API is available free of charge, according to the company’s release information. This provides developers with a way to reuse uploaded visual files across requests rather than repeatedly uploading the same image.
Why the DeepSeek V4-Flash Update Matters
The release expands DeepSeek’s multimodal capabilities beyond text processing and gives developers a lower-cost model option for applications that require visual understanding.
Potential applications include screenshot analysis, chart interpretation, image description and agent workflows involving visual information. These capabilities are supported by DeepSeek’s API documentation.
At the same time, the DeepSeek V4-Flash vision release remains experimental. Its benchmark claims should therefore be considered alongside the limitations of company-reported testing.
FAQs
What is DeepSeek V4-Flash Vision Exp?
DeepSeek-V4-Flash-Vision-Exp is an experimental multimodal model that adds image understanding to the capabilities of DeepSeek V4-Flash.
Can DeepSeek V4-Flash process images?
The new V4-Flash-Vision-Exp variant can process images alongside text. The standard V4-Flash should not be confused with this experimental vision-enabled version.
How does DeepSeek V4-Flash Vision Exp compare with Claude Opus 4.8?
DeepSeek’s own benchmark results show the experimental model ahead on some evaluations and behind on others. The results have not been independently verified.
Where can developers access the model?
The model is available through the DeepSeek API using deepseek-v4-flash-vision-exp.
What visual tasks can the model perform?
DeepSeek says the model can describe images, read text from screenshots and analyze charts, among other visual understanding tasks.
Conclusion
The DeepSeek V4-Flash family has gained a new visual capability through the experimental V4-Flash-Vision-Exp model. The release combines V4-Flash’s existing text capabilities with image understanding and makes the model available to developers through the DeepSeek API.
DeepSeek’s benchmark results indicate that the experimental model can approach Claude Opus 4.8 on selected multimodal agent evaluations, but the results vary significantly between benchmarks and have not been independently verified. That makes the release notable for its expansion into multimodal AI, while leaving broader performance comparisons open for further testing.