DeepSeek V4-Flash-Vision-Exp Launches: Multimodal API Service Goes Live
What Just Happened
DeepSeek's API platform today added a new model: V4-Flash-Vision-Exp. It is experimental, and it adds image understanding to the V4-Flash line. Access is straightforward — set model='deepseek-v4-flash-vision-exp' in your API call.
This is not a new base model. It is V4-Flash with vision layered on top. The important question is what it keeps and what it changes.
The Numbers That Matter
Text performance — Agent tasks, reasoning, world knowledge — is level with DeepSeek-V4-Flash. On this dimension nothing was sacrificed.
On Agent benchmarks that require vision, V4-Flash-Vision-Exp jumps ahead of V4-Flash. The gap is large enough that DeepSeek positions it close to Opus-4.8 in multimodal Agent capability.
In practice: the model reads images without giving up what made V4-Flash fast and cheap for text.
| Capability | V4-Flash | V4-Flash-Vision-Exp |
|---|---|---|
| Text (Agent, reasoning, knowledge) | Baseline | Matches V4-Flash |
| Multimodal Agent | Not available | Major improvement, near Opus-4.8 |
| API pricing | $0.14 / M tokens | Same as V4-Flash |
| Image token cost | — | Up to 384 tokens per image |

Why Vision Changes Agent Workflows
A text-only model can follow instructions. A vision-capable model can also look at what those instructions produce. That difference opens scenarios that were previously impractical.
Example 1: Custom PowerPoint generation in an Agent framework
The prompt asks for a one-month Tibet road-trip PPT targeting high-net-worth clients — raw, wild, explorer tone, with real photography aesthetics and three real pricing tiers. The model takes the brief, researches the route, structures the deck, generates slide content, and delivers a full presentation.
The point: from a single descriptive prompt to a finished business deliverable, inside an Agent loop.

Example 2: Website redesign
The same model takes a working website — DeepSeek Harness' own site — and reimagines it. The design brief: black-blue deep sea, glass UI, ASCII atomic pixels. After multiple rounds of refinement, the result is a futuristic developer portal.
This is the kind of task where vision matters: the model evaluates each revision against the visual brief, not just the text prompt.
Example 3: Front-end animation demo
A prompt describing "cute 3D clay monsters joining a bouncing retro dance party" becomes a working interactive mini-demo. The model writes the front-end code and iterates on the visual result until it matches the brief.

API Access
The model is available through the standard DeepSeek API. Set:
model = 'deepseek-v4-flash-vision-exp'Images are converted to tokens and billed at the V4-Flash rate. Each image costs up to 384 tokens.
Three API formats are supported: Chat Completions, Messages, and Responses. Image input accepts three methods — inline base64, external URL, and the new Files API.
Files API Also Goes Live
Alongside the vision model, DeepSeek opened a Files API. It is free to use. The workflow: upload an image once, get a file_id, then reference that ID in any number of requests. The same image does not need to be re-uploaded.
For teams calling the API at scale — Agent pipelines, batch jobs, production integrations — this cuts request payload size and removes repeated upload overhead.
What This Means
V4-Flash already had one of the best price-to-performance ratios in the market. V4-Flash-Vision-Exp keeps that pricing and adds vision on top. The combination makes it a serious option for Agent frameworks that need image understanding without a cost jump.
The experimental label is worth noting. DeepSeek is testing this model before committing to a stable release. Early adopters should monitor for changes.
Bottom Line
- Text capability: unchanged from V4-Flash
- Vision: new, strong enough to approach Opus-4.8 in Agent benchmarks
- Price: same as V4-Flash, with images costing up to 384 tokens each
- Access: standard API, plus the new Files API for image reuse
For teams running Agent workflows that involve visual output or image input, V4-Flash-Vision-Exp is worth a test run.