DeepSeek released the experimental DeepSeek-V4-Flash-Vision-Exp multimodal model on its API platform on August 21, adding image understanding to its V4-Flash model line. DeepSeek reports that the model retains V4-Flash text capabilities and brings multimodal agent results close to Anthropic's Opus-4.8 on its published benchmarks. The release also adds Files API support and DeepSeek Harness 0.1.1 compatibility.
DeepSeek released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model available through its API platform, on August 21. The model accepts mixed text and image inputs and is invoked with model='deepseek-v4-flash-vision-exp', according to DeepSeek's release documentation.
DeepSeek describes the release as retaining the text capabilities of DeepSeek-V4-Flash, including agents, reasoning, and world knowledge, while adding visual understanding for agent workflows. The company also released DeepSeek Harness 0.1.1 with built-in support for the new model.
Benchmark claims and comparisons
DeepSeek reports that V4-Flash-Vision-Exp makes a substantial improvement over V4-Flash on agent benchmarks requiring visual understanding, placing its multimodal-agent performance close to Anthropic's Opus-4.8. Its changelog lists scores including 83.9 on Terminal Bench 2.1, 57.7 on NL2Repo, 59.3 on DeepSWE, 63.6 on DSBench-Hard, 64.3 on Chartography, and 35.0 on ZeroBench Pass@5.
The Next Web's review of DeepSeek's published comparison table reports that the new model exceeds Opus-4.8 on three of 11 listed benchmarks: DeepSWE by 1.3 points, Agents' Last Exam by 1.6 points, and ZeroBench by 1.0 point. The same review reports that it trails Opus-4.8 on the other eight, including a 12-point gap on NL2Repo and an 8.1-point gap on DSBench-Hard.
Those results are vendor-published measurements. They nevertheless provide practitioners with more granularity than the headline comparison: the reported margins are close on Terminal Bench 2.1 and Chartography, while repository-level and full-stack software tasks remain materially separated in the cited table. Cross-vendor agent benchmarks can also vary with harness configuration, tool access, prompt policies, and sampling settings.
API and deployment details
The API supports Chat Completions, Messages, and Responses interfaces. DeepSeek states that images can be supplied as base64 payloads, external URLs, or Files API references. Images are tokenized for billing at up to 384 tokens per image and charged at V4-Flash pricing, according to the release.
DeepSeek's new Files API is free to use, the company states. It allows an image to be uploaded once and referenced by file_id in subsequent requests, reducing repeated request bandwidth for workflows that reuse screenshots, charts, forms, or other visual artifacts.
SiliconANGLE reports that V4-Flash-Vision-Exp is initially available through DeepSeek's paid developer platform. It also reports that the text-only V4-Flash base model is a mixture-of-experts system with 284 billion parameters, composed of experts with 13 billion parameters each. DeepSeek has not published architectural details for the vision extension, according to SiliconANGLE.
For teams building tool-using systems, the practical distinction is that visual input can be handled within the same model call pattern as text. Comparable multimodal agent deployments typically still require task-specific evaluation, particularly for screenshot interpretation, UI navigation, document extraction, and code-agent workflows where small perception errors can propagate into incorrect tool actions.
Key Points #
- 1DeepSeek's experimental release adds image input to V4-Flash while preserving the API model selection and agent-harness workflow.
- 2DeepSeek's published results place the model near Opus-4.8 on some multimodal agent tests, though the comparison comes from DeepSeek's own evaluations.
- 3Files API references can reduce repeated image-transfer overhead in multimodal applications that reuse screenshots, charts, and other visual context.
Scoring Rationale #
This is a notable multimodal API release from a major model provider, with direct relevance to teams building vision-enabled agents and tool-use workflows. The reported benchmark comparison is meaningful, but the model remains experimental and the comparison relies on DeepSeek's own published evaluations.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.