Use DeepSeek V4 Flash Vision Exp With Images

DeepSeek released deepseek-v4-flash-vision-exp on August 21, 2026. It is an experimental V4 Flash model that accepts images alongside text. DeepSeek says its text capabilities are comparable to the regular V4 Flash model, while the new model adds visual understanding for agent tasks.

The separate model ID matters. Adding an image to a request for deepseek-v4-flash does not turn that text model into the vision model.

Send a public image URL#

Create a DeepSeek API key and install a current OpenAI Python client. The following example sends an image URL through Chat Completions:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["DEEPSEEK_API_KEY"],
    base_url="https://api.deepseek.com",
)

response = client.chat.completions.create(
    model="deepseek-v4-flash-vision-exp",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": "Read this dashboard and list the visible errors. Do not infer values that are not shown.",
                },
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "https://example.com/dashboard.png"
                    },
                },
            ],
        }
    ],
)

print(response.choices[0].message.content)

Export the key before running the script:

export DEEPSEEK_API_KEY="YOUR_DEEPSEEK_API_KEY"

In PowerShell, use $env:DEEPSEEK_API_KEY = "YOUR_DEEPSEEK_API_KEY" for the current session.

Use a local image#

A local image must be encoded or uploaded before the API can receive it. For a quick test, encode it as a Base64 data URL and place that value in image_url.url. For repeated use or larger files, use DeepSeek’s Files API and reference the uploaded file according to the current vision guide.

Do not paste private screenshots into a public URL just to make the first example work. Use Base64 or the Files API when the image contains credentials, customer data, internal dashboards, or source code.

Check whether your client supports it#

The API may support a new model before desktop apps, coding agents, and gateways update their model catalogs. A client needs to do two things correctly:

1. Keep the exact model ID deepseek-v4-flash-vision-exp.
2. Send image and text blocks as a structured content array instead of flattening them into text.

If the direct API example works but your client fails, update the client and inspect its provider configuration. Look for a model-capability setting such as image input, vision, multimodal input, or array content. Do not label the normal Flash or Pro model as vision-capable to bypass a client check.

Suitable first tests#

Start with images whose answers are easy to verify: a terminal error, a small chart, a UI screenshot, or a short scanned table. Ask the model to separate visible facts from interpretation. For OCR work, compare the output with the original instead of assuming every number and symbol was read correctly.

DeepSeek V4 Flash Vision Exp adds image understanding without replacing the regular Flash and Pro IDs. Treat it as a separate experimental model, confirm that your client preserves image blocks, and keep a fallback while tool support catches up.

For text-only coding setup, see DeepSeek V4 with Claude Code and DeepSeek with Codex CLI.