DeepSeek V4 并不是一个模型。在我们的模型目录中，它是一个模型家族，包括 V4.1 Flash、V4 Pro 0813、V4 Flash 0731、V4 Flash Vision Exp、两个较早的 0423 checkpoint，以及一个 Flash“latest”别名。这些名称告诉你的是档位和发布时间，但不会告诉你哪些能接受图像。
本指南逐一介绍这个家族中的每个模型，给出我们目录中的确切 slug，并展示在图像上使用 V4 的两种方式。第一种是把图像发送给一个能读取图像的 V4 模型。第二种是在一个纯文本 V4 模型前面运行一个独立的视觉模型。
按模型给出的简短答案
有两个 V4 模型接受图像。DeepSeek V4.1 Flash 原生读取图像，是开展新图像工作的首选模型。DeepSeek V4 Flash Vision Exp 也能读取图像，DeepSeek 将其标记为实验性。
我们目录中其他所有 V4 slug 都是文本输入、文本输出。这包括 V4 Pro 0813、V4 Flash 0731、两个 0423 checkpoint，以及 ~deepseek/deepseek-v4-flash-latest 别名。向这些模型之一发送带有 image_url 部分的请求会失败，因为该模型并未将 image 列为其输入模态之一。
存在哪些 V4 模型，以及它们接受什么
该表列出了我们目录中的每个 V4 模型，并附有可回答该问题的模态列。价格是 2026 年 9 月 11 日目录中列出的每百万 token 费率，若列出的费率位数更多，则四舍五入到三位小数。
| 模型 | Slug | 输入 | 输入 / 输出价格 | 上下文 | 图像输入 |
|---|---|---|---|---|---|
| V4.1 Flash | deepseek/deepseek-v4.1-flash | 文本、图像 | $0.15 / $0.60 | 1,048,576 | 是 |
| V4 Flash Vision Exp | deepseek/deepseek-v4-flash-vision-exp | 文本、图像 | $0.22 / $0.66 | 1,048,576 | 是的，实验性 |
| V4 Pro 0813 | deepseek/deepseek-v4-pro-0813 | 文本 | $0.579 / $1.738 | 1,048,576 | 否 |
| V4 Flash 0731 | deepseek/deepseek-v4-flash-0731 | 文本 | $0.065 / $0.18 | 1,310,720 | 否 |
| V4 Pro 0423 | deepseek/deepseek-v4-pro | 文本 | $0.860 / $1.720 | 1,048,576 | 否 |
| V4 Flash 0423 | deepseek/deepseek-v4-flash | 文本 | $0.085 / $0.171 | 1,048,576 | 否 |
这六款模型都只返回文本。每个模型页面都会列出当前价格、上下文长度和输入模态，而这些值会随着提供商和 checkpoint 的变化而变化。在把某个 slug 投入生产环境之前，请先阅读该页面。
~deepseek/deepseek-v4-flash-latest 别名会重定向到 V4 Flash 系列中的最新模型。在撰写本文时，它解析为 Flash 0731，并列出文本作为其唯一的输入模态。请勿将该别名用于图像请求。请直接固定使用 deepseek/deepseek-v4.1-flash 或 deepseek/deepseek-v4-flash-vision-exp。
V4.1 Flash 原生读取图像
DeepSeek V4.1 Flash 是首个基于 DeepSeek 因果编码器-解码器架构构建的模型。它从一个 552B 参数的主干中，在输入时激活 8B 参数，在输出时激活 16B 参数。图像理解是该架构的一部分，视觉和文本嵌入向量从预训练一开始就联合训练。它在 DeepSeek 端点上拥有 1,048,576 token 的上下文窗口和 384,000 token 的最大输出，并支持工具调用、response_format 和结构化输出。
用图像调用它，与你已经发起的聊天请求相同，只需在消息内容中加入一个图像部分。TypeScript SDK 将请求体嵌套在 chatRequest 下，并使用 camelCase 字段名，因此在 SDK 中图像部分是 imageUrl，而在传输线路上是 image_url。
import { OpenRouter } from "@openrouter/sdk";
const openRouter = new OpenRouter({
apiKey: process.env.OPENROUTER_API_KEY,
});
const result = await openRouter.chat.send({
chatRequest: {
model: "deepseek/deepseek-v4.1-flash",
messages: [
{
role: "user",
content: [
{ type: "text", text: "What error state is this screenshot showing?" },
{ type: "image_url", imageUrl: { url: "https://example.com/screenshot.png" } },
],
},
],
stream: false,
},
});
if (!("choices" in result)) {
throw new Error("Expected a non-streaming response");
}
console.log(result.choices[0]?.message.content);url 字段接受公开图像 URL 或 base64 数据 URL。请先发送文本部分，再发送图像部分。图像输入指南涵盖了 base64 形式、支持的图像类型，以及在一个请求中发送多张图像。
V4 Flash Vision Exp 是实验性选项
DeepSeek V4 Flash Vision Exp 是 V4 Flash 0731 的实验性视觉增强版本。它增加了图像理解能力，同时在文本任务上与基础模型持平。它于 2026 年 8 月 21 日发布，在 2026 年 9 月 10 日 V4.1 Flash 发布之前，它是唯一接受图像的 V4 模型。
它接受与上方示例相同的请求，只需将 model 字段改为 deepseek/deepseek-v4-flash-vision-exp。它拥有 1,048,576 token 的上下文窗口，并支持工具调用、response_format 和结构化输出。
DeepSeek 将该模型标记为实验性。除非你有特定理由要测试这个实验性模型，否则新的图像工作请使用 V4.1 Flash；如果确实要测试，请固定确切的 slug。
在纯文本 V4 模型前面放置一个视觉模型
V4 Pro 0813、V4 Flash 0731 以及 0423 检查点均不接受图像。如果你想让其中之一对图像进行推理，请先运行一个视觉模型，并将其文本输出传入 V4 模型。V4 模型从不会看到像素，它读取的是描述。
有两个接受图像和视频输入的模型：Qwen3.8 27B，位于 qwen/qwen3.8-27b；以及 Kimi K3，位于 moonshotai/kimi-k3。我们的 视觉模型合集列出了所有接受图像输入的目录模型。
import { OpenRouter } from "@openrouter/sdk";
const openRouter = new OpenRouter({
apiKey: process.env.OPENROUTER_API_KEY,
});
const imageUrl = "https://example.com/screenshot.png";
const userTask = "Explain the error and propose a fix.";
// Step 1: a vision model reads the image and returns a text description.
const seen = await openRouter.chat.send({
chatRequest: {
model: "qwen/qwen3.8-27b",
messages: [
{
role: "user",
content: [
{
type: "text",
text: "Describe this screenshot for a reasoning model. Report only what is visible.",
},
{ type: "image_url", imageUrl: { url: imageUrl } },
],
},
],
stream: false,
},
});
if (!("choices" in seen)) {
throw new Error("Expected a non-streaming response");
}
const description = seen.choices[0]?.message.content;
if (typeof description !== "string") {
throw new Error("Expected a text description from the vision model");
}
// Step 2: a text-only V4 model reasons over the description.
const reasoned = await openRouter.chat.send({
chatRequest: {
model: "deepseek/deepseek-v4-pro-0813",
messages: [
{
role: "user",
content: `Image notes:\n${description}\n\nTask: ${userTask}`,
},
],
stream: false,
},
});
if (!("choices" in reasoned)) {
throw new Error("Expected a non-streaming response");
}
console.log(reasoned.choices[0]?.message.content);将 qwen/qwen3.8-27b 替换为 moonshotai/kimi-k3，即可让 Kimi K3 处理图像；或者在文本轮次中将 deepseek/deepseek-v4-pro-0813 替换为 deepseek/deepseek-v4-flash-0731。
两次调用比一次调用更贵。2026 年 9 月 11 日，Qwen3.8 27B 的标价为每百万输入 token $0.42、每百万输出 token $3.00，Kimi K3 的标价为 $2.10 和 $10.53。你还要在此基础上为 V4 模型付费。当你需要纯文本的 V4 模型或视频输入时，才使用这种模式，不要把它当作使用 V4 读取图像的默认方式。
该使用哪条路径
根据你需要的模型和你手头的输入来匹配路径。
- 一张静态图像或文档，且 Flash 级别的质量就足够。 在一次调用中将其发送给
deepseek/deepseek-v4.1-flash。 - 一张静态图像，并且你想测试这个实验性模型。 在一次调用中将其发送给
deepseek/deepseek-v4-flash-vision-exp，并固定该 slug。 - 在图像上使用 Pro 级推理。没有任何 V4 Pro 模型接受图像。请先运行一个视觉模型，并将其文本传入
deepseek/deepseek-v4-pro-0813。 - 视频输入。没有任何 V4 模型接受视频。请运行一个支持视频的模型，例如 Qwen3.8 27B 或 Kimi K3，并将其文本传入 V4 模型。视频输入指南介绍了请求的结构。
- 文本与图像混合流量。将纯文本请求保留在你已在使用的文本模型上，并将图像请求发送至 V4.1 Flash。
如果图像在毫无预警的情况下到来，请检查你自己代码中的消息内容，并在发送请求前选定模型。提供商路由会在你所指定模型的各提供商之间进行选择。它不会切换到其他模型，因此无法将发往纯文本模型的图像请求转变为发往视觉模型的请求。
如何自行检查模型的模态
模型页面和 models API 是某个 slug 接受何种输入的权威来源。每个模型页面都会列出其输入和输出模态。models API 会在每个模型上返回与 architecture.input_modalities 相同的数据。
curl -s https://openrouter.ai/api/v1/models \
| jq -r '.data[] | select(.id | startswith("deepseek/deepseek-v4")) | "\(.id)\t\(.architecture.input_modalities | join(","))"'models 页面筛选图像输入后，会显示我们目录中所有接受图像的模型。如果后续的 V4 checkpoint 增加了图像输入，其模型页面和 API 响应都会显示出来，我们也会更新此页面。
常见问题
DeepSeek V4 Flash 支持视觉吗？
这取决于具体的 checkpoint。deepseek/deepseek-v4.1-flash 和 deepseek/deepseek-v4-flash-vision-exp 接受文本和图像。deepseek/deepseek-v4-flash-0731 和 deepseek/deepseek-v4-flash 仅支持文本，而 ~deepseek/deepseek-v4-flash-latest 别名在撰写本文时指向 Flash 0731。
DeepSeek V4 Pro 支持视觉吗？
不支持。deepseek/deepseek-v4-pro-0813 和更早的 deepseek/deepseek-v4-pro 都将文本列为唯一的输入模态。要在图像上使用 Pro，请先运行一个视觉模型，再将其文本描述传入 Pro。
处理图像应该用哪个 DeepSeek V4 模型？
使用 deepseek/deepseek-v4.1-flash。它能原生读取图像，未被标记为实验性，并且在 2026 年 9 月 11 日的标价为每百万输入 token $0.15、每百万输出 token $0.60。deepseek/deepseek-v4-flash-vision-exp 是实验性的替代选项。
DeepSeek V4 的图像输入费用是多少？
图像输入按模型的 token 价格计费。2026 年 9 月 11 日，V4.1 Flash 标价为每百万 token 输入 $0.15、输出 $0.60，V4 Flash Vision Exp 标价为输入 $0.22、输出 $0.66。在制定预算之前请查看模型页面，因为标价会变动。
DeepSeek V4 能处理视频吗？
我们的目录中没有任何 V4 模型将视频列为输入模态。如需处理视频，请使用支持视频的模型，例如 qwen/qwen3.8-27b 或 moonshotai/kimi-k3，然后将其文本输出传入 V4 模型。
DeepSeek V4 is not one model. On our catalog it is a family that includes V4.1 Flash, V4 Pro 0813, V4 Flash 0731, V4 Flash Vision Exp, two older 0423 checkpoints, and a Flash “latest” alias. The names tell you the tier and the release. They do not tell you which ones accept an image.
This guide goes through the family one model at a time, with the exact slugs from our catalog, and shows two ways to use V4 on an image. The first sends the image to a V4 model that reads images. The second runs a separate vision model in front of a text-only V4 model.
The short answer, by model
Two V4 models accept images. DeepSeek V4.1 Flash reads images natively and is the model to use for new image work. DeepSeek V4 Flash Vision Exp also reads images, and DeepSeek labels it experimental.
Every other V4 slug on our catalog is text in, text out. That includes V4 Pro 0813, V4 Flash 0731, the two 0423 checkpoints, and the ~deepseek/deepseek-v4-flash-latest alias. A request that sends an image_url part to one of these models fails, because the model does not list image among its input modalities.
Which V4 models exist and what they accept
The table lists each V4 model on our catalog with the modality column that answers the question. Prices are the listed catalog rates per million tokens on 11 September 2026, rounded to three decimal places where a listed rate has more digits.
| Model | Slug | Input | Price in / out | Context | Image input |
|---|---|---|---|---|---|
| V4.1 Flash | deepseek/deepseek-v4.1-flash | text, image | $0.15 / $0.60 | 1,048,576 | Yes |
| V4 Flash Vision Exp | deepseek/deepseek-v4-flash-vision-exp | text, image | $0.22 / $0.66 | 1,048,576 | Yes, experimental |
| V4 Pro 0813 | deepseek/deepseek-v4-pro-0813 | text | $0.579 / $1.738 | 1,048,576 | No |
| V4 Flash 0731 | deepseek/deepseek-v4-flash-0731 | text | $0.065 / $0.18 | 1,310,720 | No |
| V4 Pro 0423 | deepseek/deepseek-v4-pro | text | $0.860 / $1.720 | 1,048,576 | No |
| V4 Flash 0423 | deepseek/deepseek-v4-flash | text | $0.085 / $0.171 | 1,048,576 | No |
All six return text only. Each model page lists the current price, context length, and input modalities, and those values change as providers and checkpoints change. Read the page before you commit a slug to production.
The ~deepseek/deepseek-v4-flash-latest alias redirects to the latest model in the V4 Flash family. At the time of writing it resolved to Flash 0731 and listed text as its only input modality. Do not use the alias for image requests. Pin deepseek/deepseek-v4.1-flash or deepseek/deepseek-v4-flash-vision-exp directly.
V4.1 Flash reads images natively
DeepSeek V4.1 Flash is the first model built on DeepSeek’s Causal Encoder-Decoder architecture. It activates 8B parameters on input and 16B on output from a 552B-parameter backbone. Image understanding is part of the architecture, with visual and text embeddings trained jointly from the start of pre-training. It has a 1,048,576-token context window and a 384,000-token maximum output on the DeepSeek endpoint, and it supports tool calling, response_format, and structured outputs.
Calling it with an image is the same chat request you already make, with an image part added to the message content. The TypeScript SDK nests the request body under chatRequest and uses camelCase field names, so the image part is imageUrl in the SDK and image_url on the wire.
import { OpenRouter } from "@openrouter/sdk";
const openRouter = new OpenRouter({
apiKey: process.env.OPENROUTER_API_KEY,
});
const result = await openRouter.chat.send({
chatRequest: {
model: "deepseek/deepseek-v4.1-flash",
messages: [
{
role: "user",
content: [
{ type: "text", text: "What error state is this screenshot showing?" },
{ type: "image_url", imageUrl: { url: "https://example.com/screenshot.png" } },
],
},
],
stream: false,
},
});
if (!("choices" in result)) {
throw new Error("Expected a non-streaming response");
}
console.log(result.choices[0]?.message.content);The url field accepts a public image URL or a base64 data URL. Send the text part before the image part. The image inputs guide covers the base64 form, supported image types, and sending several images in one request.
V4 Flash Vision Exp is the experimental option
DeepSeek V4 Flash Vision Exp is an experimental vision-enabled version of V4 Flash 0731. It adds image understanding while matching the base model on text tasks. It shipped on 21 August 2026, and until V4.1 Flash shipped on 10 September 2026 it was the only V4 model that accepted images.
It takes the same request as the example above with the model field changed to deepseek/deepseek-v4-flash-vision-exp. It has a 1,048,576-token context window and supports tool calling, response_format, and structured outputs.
DeepSeek labels this model experimental. Use V4.1 Flash for new image work unless you have a specific reason to test the experimental model, and pin the exact slug if you do.
Put a vision model in front of a text-only V4 model
V4 Pro 0813, V4 Flash 0731, and the 0423 checkpoints do not accept images. If you want one of them to reason over an image, run a vision model first and pass its text output into the V4 model. The V4 model never sees the pixels. It reads the description.
Two models that accept image and video input are Qwen3.8 27B at qwen/qwen3.8-27b and Kimi K3 at moonshotai/kimi-k3. Our vision models collection lists every catalog model that accepts image input.
import { OpenRouter } from "@openrouter/sdk";
const openRouter = new OpenRouter({
apiKey: process.env.OPENROUTER_API_KEY,
});
const imageUrl = "https://example.com/screenshot.png";
const userTask = "Explain the error and propose a fix.";
// Step 1: a vision model reads the image and returns a text description.
const seen = await openRouter.chat.send({
chatRequest: {
model: "qwen/qwen3.8-27b",
messages: [
{
role: "user",
content: [
{
type: "text",
text: "Describe this screenshot for a reasoning model. Report only what is visible.",
},
{ type: "image_url", imageUrl: { url: imageUrl } },
],
},
],
stream: false,
},
});
if (!("choices" in seen)) {
throw new Error("Expected a non-streaming response");
}
const description = seen.choices[0]?.message.content;
if (typeof description !== "string") {
throw new Error("Expected a text description from the vision model");
}
// Step 2: a text-only V4 model reasons over the description.
const reasoned = await openRouter.chat.send({
chatRequest: {
model: "deepseek/deepseek-v4-pro-0813",
messages: [
{
role: "user",
content: `Image notes:\n${description}\n\nTask: ${userTask}`,
},
],
stream: false,
},
});
if (!("choices" in reasoned)) {
throw new Error("Expected a non-streaming response");
}
console.log(reasoned.choices[0]?.message.content);Swap qwen/qwen3.8-27b for moonshotai/kimi-k3 to put Kimi K3 on the image, or swap deepseek/deepseek-v4-pro-0813 for deepseek/deepseek-v4-flash-0731 on the text turn.
Two calls cost more than one. On 11 September 2026, Qwen3.8 27B listed at $0.42 per million input tokens and $3.00 per million output tokens, and Kimi K3 listed at $2.10 and $10.53. You pay the V4 model on top of that. Use this pattern when you need a text-only V4 model or video input, not as the default way to read an image with V4.
Which route to use
Match the route to the model you need and the input you have.
- A still image or document, and Flash-tier quality is enough. Send it to
deepseek/deepseek-v4.1-flashin one call. - A still image, and you want to test the experimental model. Send it to
deepseek/deepseek-v4-flash-vision-expin one call, and pin the slug. - Pro-tier reasoning on an image. No V4 Pro model accepts images. Run a vision model first and pass its text into
deepseek/deepseek-v4-pro-0813. - Video input. No V4 model accepts video. Run a model that does, such as Qwen3.8 27B or Kimi K3, and pass its text into a V4 model. The video inputs guide covers the request shape.
- Mixed text and image traffic. Keep text-only requests on the text model you already use and send image requests to V4.1 Flash.
If images arrive without warning, check the message content in your own code and pick the model before you send the request. Provider routing chooses between providers of the model you named. It does not switch to a different model, so it cannot turn an image request to a text-only model into a request to a vision model.
How to check a model’s modalities yourself
The model page and the models API are the source of truth for what a slug accepts. Each model page lists its input and output modalities. The models API returns the same data as architecture.input_modalities on every model.
curl -s https://openrouter.ai/api/v1/models \
| jq -r '.data[] | select(.id | startswith("deepseek/deepseek-v4")) | "\(.id)\t\(.architecture.input_modalities | join(","))"'The models page filtered to image input shows every model on our catalog that accepts images. If a later V4 checkpoint adds image input, its model page and the API response will show it, and we will update this page.
Frequently asked questions
Does DeepSeek V4 Flash have vision?
It depends on the checkpoint. deepseek/deepseek-v4.1-flash and deepseek/deepseek-v4-flash-vision-exp accept text and images. deepseek/deepseek-v4-flash-0731 and deepseek/deepseek-v4-flash are text-only, and the ~deepseek/deepseek-v4-flash-latest alias resolved to Flash 0731 at the time of writing.
Does DeepSeek V4 Pro have vision?
No. deepseek/deepseek-v4-pro-0813 and the older deepseek/deepseek-v4-pro list text as their only input modality. To use Pro on an image, run a vision model first and pass its text description into Pro.
Which DeepSeek V4 model should I use for images?
Use deepseek/deepseek-v4.1-flash. It reads images natively, it is not marked experimental, and it listed at $0.15 per million input tokens and $0.60 per million output tokens on 11 September 2026. deepseek/deepseek-v4-flash-vision-exp is the experimental alternative.
How much does DeepSeek V4 image input cost?
Image input is billed through the model’s token prices. On 11 September 2026, V4.1 Flash listed at $0.15 in and $0.60 out per million tokens, and V4 Flash Vision Exp listed at $0.22 in and $0.66 out. Check the model page before you size a budget, because listed prices change.
Can DeepSeek V4 handle video?
No V4 model on our catalog lists video as an input modality. For video, use a model that does, such as qwen/qwen3.8-27b or moonshotai/kimi-k3, then pass its text output into a V4 model.