DeepSeek V4 Flash Vision Exp: Experimental Vision, Limits, and How to Run It in OpenCode

DeepSeek gave V4 Flash eyes. The experimental model id is deepseek-v4-flash-vision-exp, the docs went up on August 21, 2026, and the Hacker News thread filled in within hours. This is not a new flagship. It is Flash with image input, and DeepSeek's release note says the text side (agents, reasoning, world knowledge) matches V4 Flash 0731.

Screenshot-in-the-loop work no longer needs a second vision model bolted on. The catch is the contract: every image is resized toward an 800x800 pixel budget, billed at most 384 tokens, and rejected if you send it to any other DeepSeek model. Here is the official limit sheet and the fastest way to try it: through OpenCode.

Official Sources#

What Shipped#

The model accepts images alongside text so you can describe pictures, read screenshots, and analyze charts. DeepSeek is explicit that this is experimental. Use deepseek-v4-flash-vision-exp. Do not send images to deepseek-v4-flash or deepseek-v4-pro: those return a 400 ("This model does not support image").

Three API surfaces, same model:

• OpenAI-compatible Chat Completions at https://api.deepseek.com
• Anthropic-compatible /messages at https://api.deepseek.com/anthropic
• OpenAI-compatible Responses API, where images travel in input_image parts

The V4 developer guide still covers auth, caching, and the rest of the family. This post is only the vision add-on.

DeepSeek's release note claims a "major leap" on multimodal agent benchmarks versus text-only Flash, "bringing multimodal agent performance close to Opus-4.8." That is vendor framing. The vision guide does not publish a score table, and we are not reading numbers off the announcement chart.

Image Contract#

Supported formats: JPEG, PNG, GIF, and WebP. Format is detected from file content, not from the file name or the declared MIME type. That is a useful detail if your pipeline lies about Content-Type.

Three ways to send an image:

1. Base64 data URL. Inline data:image/jpeg;base64,... in the request. Simplest for local files. The encoded bytes count toward the 48 MiB request body limit.
2. External https URL. DeepSeek downloads it for you. The URL may be at most 8192 characters, the file at most 32 MiB, and the download must finish in 60 seconds.
3. Files API file_id. Upload once, reuse the id. Images referenced this way may be up to 64 MiB and skip the 32 MiB per-image check. The release note says the Files API is free to use.

Use Files API when a request would blow the 48 MiB body, when the image is larger than 32 MiB, or when you reuse the same screenshot across turns.

Hard limits from the vision guide:

Images belong in user messages only. Images in system or assistant messages return 400. User text that contains the reserved image placeholder token is also a 400.

For image_url inputs you can set detail:

"Keeps the original" is the field's wording. Inference still runs the resize described next.

Token Usage, Not Magic Pixels#

Images become tokens and those tokens are billed with the text. Before inference, every image is resized while preserving aspect ratio. Below roughly 384x384 pixels it is scaled up. Larger images are scaled down so the pixel count is roughly that of an 800x800 image.

Upper bound: 384 tokens per image. A 2000x2000 image and a 5000x5000 image cost the same after resize. Multi-image requests count each image independently.

That 800x800 budget is the honest constraint. A full-page screenshot or a schematic with small labels will lose detail. Crop the region you care about before you send it. detail: low is a further downscale to 512x512, so use it when you only need "what is on this screen," not "read the 11px caption."

The vision guide does not list dollar prices. DeepSeek's release note only says images are billed "at V4-Flash pricing" with the 384-token cap. For a number you can budget against today, the OpenCode catalog (opencode models --verbose, fetched August 21, 2026) lists opencode-go/deepseek-v4-flash-vision-exp at:

Context is 1,000,000 tokens with 384,000 max output. Family is deepseek-flash. Text-only opencode-go/deepseek-v4-flash is the same catalog rate; the vision build adds attachment: true and image input. Those are OpenCode's numbers, not DeepSeek's first-party sheet.

Running It in OpenCode#

Install from the OpenCode docs:

curl -fsSL https://opencode.ai/install | bash

The model is opencode-go/deepseek-v4-flash-vision-exp. Variants are low, high, and max:

# One-shot at max reasoning
opencode run --model opencode-go/deepseek-v4-flash-vision-exp --variant max \
  "inspect the screenshot at ./ui.png and list the layout bugs"

# Interactive session with the model preselected
opencode --model opencode-go/deepseek-v4-flash-vision-exp

max is for multi-step loops: plan, look at a screenshot, edit, look again. high is enough for a single screenshot question. low is the cost lever when you are triaging a pile of images.

Same OpenCode path as Ox Alpha and cheap GLM-5.3 access. If you already have Go credits, you do not need a second account.

Call It Yourself (OpenAI-compatible)#

Same shape as the official vision guide. Base64 inline is the local-file path:

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-v4-flash-vision-exp",
    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)

Swap the data URL for "url": "https://example.com/image.jpg" to use method 2, or a {"type": "file", "file_id": "file-api-..."} block for method 3.

Anthropic /messages uses an image block with source.type of base64, url, or file instead of image_url. Responses API uses input_image. Limits are the same across all three.

When to Reach for It#

Use deepseek-v4-flash-vision-exp when:

• The agent needs to see a screenshot, chart, or mockup, not just source
• You already like Flash for agent loops and want image input in the same family
• You can crop dense images so the 800x800 resize still leaves the details you care about

That last point is the same idea as Kimi K3's vision-in-the-loop websites: the screenshot is an evaluation surface, not a substitute for tests. Vision will not tell you whether a button has the right type.

Stay on text-only Flash when the task is code and logs with no pixels, or you do not want an experimental (-exp) id in the path. Sending images to deepseek-v4-flash is a 400. OpenCode prices the two Flash ids the same today, so the reason to keep the text-only id is caution, not catalog cost.

FAQ#

What is deepseek-v4-flash-vision-exp?#

An experimental multimodal API model DeepSeek shipped on August 21, 2026. It accepts images with text. DeepSeek says its text capabilities match V4 Flash. Other DeepSeek models reject images with HTTP 400.

How do I run it in OpenCode?#

Install OpenCode (curl -fsSL https://opencode.ai/install | bash), then opencode run --model opencode-go/deepseek-v4-flash-vision-exp --variant max "your task". Variants are low, high, and max. Family is deepseek-flash. The vision id is on OpenCode Go.

What image formats and size limits apply?#

JPEG, PNG, GIF, and WebP, detected from file bytes. Request body 48 MiB. Up to 600 images per request. Max 8192 px per side, or 4096 px per side at 15 or more images. External URLs: 8192 characters, 32 MiB, 60 second download. Files API file_id images: up to 64 MiB.

How are images billed?#

The vision guide does not publish dollar rates. Images are resized toward an 800x800 pixel count and billed with text, capped at 384 tokens per image. OpenCode's catalog on August 21, 2026 lists $0.22 input / $0.66 output per million tokens for opencode-go/deepseek-v4-flash-vision-exp.

Can I put images in system prompts or tool results via Chat Completions?#

Not in system or assistant messages: those 400. Images go in user messages. The Responses API also allows input_image on function_call_output / custom_tool_call_output items. If your harness only speaks Chat Completions, keep screenshots on the user turn.

Sources#

Last updated: August 21, 2026

Some links to tools above are referral links - see our affiliate disclosure.

Continue Reading#

• DeepSeek V4 Flash 0731: Official Release and OpenCode Setup - the text-only Flash release this vision build sits on
• DeepSeek V4 Developer Guide - family, caching, and first-party API wiring
• Kimi K3 Websites: What Vision in the Loop Actually Means - screenshot feedback as a frontend-agent loop, and what it cannot verify
• Ox Alpha on OpenCode - another OpenCode-catalog model, with the same install path
• Where to Run GLM-5.3 Free and Cheap - OpenCode Go as a cheap coding-model lane