DeepSeek 发布多模态视觉模型 deepseek-v4-flash-vision-exp

DeepSeek 推出新型多模态视觉理解模型 deepseek-v4-flash-vision-exp，已在 DeepSeek API 平台上线2。该模型支持图像和文本的多模态输入1，可处理 JPEG、PNG、GIF 和 WebP 格式的图像12。用户可通过设置 model='deepseek-v4-flash-vision-exp' 来访问该模型2。

该模型提供三种图像输入方式：Base64 编码、外部 URL 和 Files API 上传1。其中外部图像 URL 的文件大小上限为 32 MiB，下载需在 60 秒内完成1；Files API 上传的图像文件最大可达 64 MiB1，且该功能已开放且不收费2。单个请求体大小限制为 48 MiB1。每张图像的令牌消耗上限为 384 tokens1，该模型的图片计费标准与 V4-Flash 模型保持一致2。

模型支持 Chat Completions、Messages 和 Responses 三种调用格式2。在图像处理方面，当图像像素总数低于 384×384 时自动放大，超过时则缩小至 800×800 像素1。使用中需注意的是，图像仅支持在 user 消息中使用，在 system 或 assistant 消息中使用将返回 400 错误1。此外，外部图像 URL 的字符长度上限为 8192 字符1。该模型在 Agent 框架中展现了较好的适配能力，多模态 Agent 能力已接近 Opus-4.8 水平2。

DeepSeek has released the DeepSeek-v4-flash-vision-exp model, a multimodal vision understanding system now available on the DeepSeek API platform.2 The model supports image and text inputs, enabling users to process visual content alongside textual queries.1

The model accepts images in JPEG, PNG, GIF, and WebP formats through three input methods: Base64 encoding, external URLs, and Files API uploads.12 Images can be provided via external URLs up to 8,192 characters in length, with files reaching a maximum size of 32 MiB that must be downloaded within 60 seconds.1 Files API uploads support images up to 64 MiB in size.1 The total request body is limited to 48 MiB.1 Each image is allocated up to 384 tokens for processing, with pricing consistent with the V4-Flash model.2 Images undergo automatic resizing: those with fewer total pixels than a 384×384 resolution are enlarged, while larger images are reduced to 800×800 pixels.1

The API is compatible with both OpenAI and Anthropic API standards,1 and users can access the model by setting the parameter model='deepseek-v4-flash-vision-exp'.2 The model supports Chat Completions, Messages, and Responses calling formats.2 An important restriction applies: images are only permitted in user messages; attempts to include images in system or assistant messages will return a 400 error.1 The Files API for image uploads is now available and does not incur additional charges.2