DeepSeek-v4-flash-vision-exp

TL;DR Highlight

deepseek-flash now accepts images, with a 1024-token cap per image

Who Should Read

A backend developer using text-only DeepSeek API for code or agents who now needs screenshot, chart, or OCR input

Core Mechanics

deepseek-flash accepts images
- You can pair images with text to describe pictures, read screenshot text, and analyze charts.
- The old name deepseek-v4-flash-vision-exp is retired, and its requests are served by the latest Flash model.
Three ways to send an image
- Choose a base64 data URL, an external http(s) URL, or a Files API file_id.
- It uses the OpenAI-compatible Chat Completions format with content as an array of blocks.
Image token cap is 1024
- Images under about 544×544 are scaled up, and larger ones are scaled down to about 1300×1300, keeping aspect ratio.
- So a 2000×2000 and a 5000×5000 image use the same tokens.
The detail field controls resolution
- low downscales to 512×512 for faster, cheaper calls.
- high, original, and auto all currently keep the original.
Limits are specific

Community Reactions

Hope for screenshot analysis
- One commenter said not viewing Playwright screenshots precisely was the one thing they missed from Sonnet.
- Another said agent use is limited since tool results cannot carry images.
Accuracy still looks weak
- On a clock image whose answer is 08:09:25, DeepSeek said 5:10:45 while Qwen3.8 27B said 8:10:25, per one commenter.
- On a 12-image landmark test, ByteDance seed 2.1 turbo got 11 and DeepSeek got 6, per one commenter.
Resolution seen as too low
- One commenter said about 800×800 is too low for OCR of a full A4 page.
- Another said around 1080p is needed to read small text.
Earlier version misbehaved with images
- A commenter said version 0731 assumed it had vision and tried to invent image analysis tools.
- Another had to tell it not to view screenshots because it kept breaking sessions.
Workaround setup shared

How to Apply

Send one image to the API
- Today, set base_url to https://api.deepseek.com and model to deepseek-flash, then send a screenshot.
- Check accuracy on your usual OCR or chart samples first.
Use detail low where cost matters
- Set detail to low in image_url for classification or simple description requests.
- Compare tokens and answer quality against original on the same image.
Use the Files API for repeated images
- Upload once and put the file_id in a file block when reusing an image across requests.
- Send images over 32 MiB or requests over 48 MiB this way.
Test OCR and agents against alternatives
- Crop or tile full-page documents to check for resolution loss.
- If you need images in tool results, first check Responses API support for function_call_output images.

Try It

copy & run

# Run in a terminal (needs pip install openai; change only the image path and API key)

import base64
from openai import OpenAI

client = OpenAI(api_key="{{DeepSeek API Key}}", base_url="https://api.deepseek.com")
with open("{{image.jpg}}", "rb") as f:
    b64 = base64.b64encode(f.read()).decode("utf-8")

for detail in ["low", "original"]:
    r = client.chat.completions.create(
        model="deepseek-flash",
        messages=[{"role": "user", "content": [
            {"type": "text", "text": "{{question}}"},
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64}", "detail": detail}},
        ]}],
    )
    print(detail, r.usage.prompt_tokens)
    print(r.choices[0].message.content)

# Next: compare token counts and answers between low and original on the same image.

Terminology

base64 data URLEmbeds the image file as a text string in the request body. Like unpacking a parcel and taping its contents to the letter.

Files APIUpload a file first and refer to it by file_id. Like leaving luggage at storage and showing only the claim ticket.

detail fieldAn option that sets how sharply the image is processed. low downsizes to 512×512 for cheaper, faster calls.

Responses APIAnother OpenAI-compatible API for generating responses. Images go in input_image parts.