DeepSeek Vision: Use deepseek-flash, the Experimental Vision Model Is Retired
Rechecked on 2026-10-18 against the official Vision guide and pricing page. Image input goes through deepseek-flash (V4.1-Flash-1017). deepseek-v4-flash-vision-exp is retired. The name is still accepted and served by the latest Flash. V4-Pro has no vision.
Checked on 2026-10-18. Pages opened that day:
An older draft treated deepseek-v4-flash-vision-exp as the current model, and said the V4 family had three public models and that Flash had been updated to V4-Flash-0731. Those three claims are out of date.
How image input works now
On the pricing page, deepseek-flash has vision checked. The version is DeepSeek-V4.1-Flash-1017. deepseek-v4-pro (V4-Pro-0813) says Not supported.
The vision guide opens with the same fact: deepseek-flash accepts images next to text. The old name deepseek-v4-flash-vision-exp is still accepted. That model is retired, and those requests are also served by the latest Flash, at the Flash price.
The guide’s examples are describing a picture, reading text in a screenshot, and analyzing a chart. Formats: JPEG, PNG, GIF, and WebP. The format is detected from the file bytes, not the file name or the MIME type.
Images are allowed in user messages only. An image in system or assistant returns 400.
Image tokens are billed with text tokens at the Flash peak and off-peak rates. Off-peak / peak, per 1M tokens: cache hit $0.003 / $0.006, cache miss $0.15 / $0.30, output $0.60 / $1.20. Peak hours are in the API platform guide.
The experimental model, as history
On 2026-08-21 the changelog released DeepSeek-V4-Flash-Vision-Exp as deepseek-v4-flash-vision-exp. That was the first time the V4 series exposed image input on the API. It was a separate model from text Flash.
On 2026-09-10, V4-Flash and V4-Flash-Vision-Exp were retired. deepseek-v4-flash and deepseek-v4-flash-vision-exp were temporarily routed to V4.1-Flash. After the 17 October upgrade, the pricing page says both old names are served by V4.1-Flash-1017 at the Flash price.
The V4.1-Flash news post calls it the smallest model in the new architecture family, with native visual understanding, not an experimental head added onto a text-only Flash. “First vision model” fits the 21 August experiment. It does not fit today’s deepseek-flash.
The older draft said the experiment hit 458 points on Hacker News that day. That thread was not reopened on 2026-10-18, so the number is not repeated as a current fact.
Three ways to send an image
All three use OpenAI Chat Completions, with content as an array of blocks. The Responses API uses input_image. The Anthropic-compatible base URL is https://api.deepseek.com/anthropic, and the image block type is image.
- Base64. Put a
data:image/jpeg;base64,...URL in the request. This is the simple path for a local file. The encoded bytes count toward the 48 MiB body limit. - External URL. A public http(s) link. The URL may be at most 8192 characters, the file at most 32 MiB, and the download must finish within 60 seconds.
- Files API
file_id. Use this when the same image is sent many times, or when one image breaks the inline limit. Afile_idimage may be up to 64 MiB and skips the 32 MiB check. Afileblock can also carryfile_dataas base64.file_idandfile_dataare mutually exclusive.
image_url accepts detail. low downscales to 512×512. high and original keep the image. auto currently matches original. detail is ignored for file_id.
Before inference, every image is resized and the aspect ratio is kept. Images under roughly 544×544 pixels are scaled up. Larger images are scaled down to about the pixel count of a 1300×1300 image. That caps one image at 1024 tokens. A 2000×2000 image and a 5000×5000 image cost the same number of tokens after resizing. Multiple images are counted one by one. The token calculator on DeepSeek’s Token page was not used here, so no calculator output is quoted.
The older draft said images under about 384×384 were scaled up. The vision guide on 2026-10-18 says roughly 544×544.
| Limit | Value on the guide |
|---|---|
| Formats | JPEG, PNG, GIF, WebP |
| External URL length | 8192 characters |
| Request body | 48 MiB |
| One image (base64 or external URL) | 32 MiB |
One image (Files API file_id) | 64 MiB |
| Images per request | 600 |
| Total image bytes per request | 64 MiB without file_id images; up to 200 MiB including file_id images |
| Max side | 8192 px; 4096 px when a request has 15 or more images |
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")
response = client.chat.completions.create(
model="deepseek-flash",
messages=[
{
"role": "user",
"content": [
{"type": "text", "text": "What is in this image?"},
{
"type": "image_url",
"image_url": {"url": f"data:image/jpeg;base64,{b64}"},
},
],
}
],
)
print(response.choices[0].message.content)
If a tool such as Claude Code or OpenCode only needs a new base_url and model name, use deepseek-flash. The changelog recommends DeepSeek Harness 0.2.1-rc.1 or later. This page does not retest each client’s default model ID.
📚 Keep Reading
📚 Related Guides
DeepSeek API Model IDs and Where to Call Them (2026-10-18)
Rechecked on 2026-10-18. Call deepseek-flash (V4.1-Flash-1017) or deepseek-v4-pro (V4-Pro-0813). The old Flash names still work, but those models are retired. Includes peak pricing, the 1M context, and the OpenRouter, Fireworks, and Together pages opened that day.
China Coding API Roundup 2026: Ark vs Bailian vs MiniMax vs Zhipu vs DeepSeek
The most comprehensive 2026 comparison of Chinese AI coding API subscriptions: Volcengine Ark Lite ¥40/mo (¥9.9 first month), Alibaba Bailian ¥200, MiniMax ¥29, Zhipu GLM ¥49+, and DeepSeek pay-as-you-go. One big table, six scenarios, and a decision tree to find your fit in 30 seconds.
Why did DeepSeek miss the flowchart in your PDF requirements?
Check the parsing purpose and figure extraction before changing models. A documented workflow for PDF uploads, file references, timeout recovery and separate parsing/inference charges.
📖 Related Comparisons
Bailian Coding Plan vs Volcengine Ark Coding Plan: 2026 Deep Comparison
A comprehensive comparison of Alibaba Cloud Bailian and Volcengine Ark — two major China-native AI coding API plans: pricing, quota mechanisms, model ecosystems, client compatibility, and selection advice.
5 China-Native AI Coding Subscriptions Compared: August 2026 Guide
An August 2026 comparison of Volcengine Ark Coding Plan, Alibaba Bailian Token Plan, MiniMax Token Plan, GLM Coding Plan, and Kimi Code: price, quota, models, API compatibility, and best-fit scenarios.
Coding Plan vs Agent Plan vs Token Plan: 2026 China AI Subscription Showdown
A full comparison of Ark Agent Plan, MiniMax Token Plan, and Bailian Token Plan (Team) — pricing, billing units, model lineups, modality coverage, Harness tools, and how to choose.
🎁 Grab deals & free trials
Free quotas, first-month discounts, and sign-up links for tools & coding plans — kept up to date.
View all deals →