Does DeepSeek V4 support image input?

Not in the public API today. DeepSeek V4 Pro and V4 Flash are text-only through the official API. DeepSeek has published vision research (the Thinking with Visual Primitives report), but there is no generally available image endpoint or public vision model ID.

How can I send images to DeepSeek then?

Only through a proxy workflow: a separate vision model reads the image and forwards a text description to DeepSeek. That is not native V4 image understanding.

Will DeepSeek release a vision API?

DeepSeek's Visual Primitives release describes future work, including eventual integration of the model weights into a foundation model. Until DeepSeek publishes the model, request schema, pricing, and support boundary, treat native vision as not generally available.

DeepSeek Vision: Does V4 Support Image Input?

DeepSeek V4 is text-only in the public API today. The Visual Primitives report shows DeepSeek's vision research, but not a generally available vision endpoint.

Status verified 2026-07-14: the public DeepSeek V4 API is still text-only. This page is kept current as the DeepSeek vision situation changes.

Direct answer: public DeepSeek V4 is text-only today

DeepSeek V4 Pro and V4 Flash do not currently expose image input through the public DeepSeek API. The official model and pricing documentation lists text-generation features such as thinking modes, JSON output, tool calls, prefix completion, and FIM completion, but it does not document an image request format or a public vision model ID.

DeepSeek's own GitHub Copilot integration guide makes the boundary explicit: V4 is text-only, and screenshot support in that extension works by sending the image to another installed vision model before forwarding a text description to DeepSeek. That is a proxy workflow, not native V4 image understanding.

What DeepSeek actually released

On April 30, 2026, DeepSeek published the Thinking with Visual Primitives technical report and its official research repository. The release describes a multimodal research model and promises future publication of internal benchmarks, part of the cold-start dataset, and eventual integration of the model weights into a foundation model.

That wording matters. The official release establishes a real DeepSeek vision research direction, but it does not announce a generally available API endpoint, public product rollout, or downloadable production model. A research report and a public image API are different release milestones.

How visual primitives work

The report argues that stronger image encoders do not solve every visual reasoning failure. A model can perceive objects yet still lose track of which exact object, path, or location its language refers to. DeepSeek calls this the Reference Gap.

The proposed system adds two spatial markers to the reasoning path:

Points identify locations, paths, and topological relationships.
Bounding boxes identify objects together with position and scale.

Instead of producing coordinates only as a final detection result, the model can interleave those markers with its reasoning. The goal is to keep multi-step answers anchored to the image while counting objects, tracing a maze, reading a diagram, or referring to a precise region.

The research stack uses DeepSeek V4 Flash as its language-model foundation together with an in-house DeepSeek-ViT encoder. The report also describes aggressive visual-token and KV-cache compression, but those research details do not change the current public API contract.

Current availability at a glance

DeepSeek V4 text API: available as V4 Pro and V4 Flash.
Thinking with Visual Primitives report: publicly available from DeepSeek.
Native V4 image input in the public API: not documented or generally available.
Public Visual Primitives model weights: described by the release as future work.
Image handling through a third-party tool: possible when that tool uses a separate vision model or proxy.

This status does not mean DeepSeek lacks vision research. It means developers should not turn a research repository into a product claim before DeepSeek publishes the model, request schema, pricing, and support boundary.

What developers should do now

Do not send image blocks to V4 Pro or V4 Flash based only on the paper, and do not invent model names such as deepseek-v4-vision. Code against the model IDs and input fields listed in the official API documentation.

If an application needs screenshots or document images today, use a separately documented vision model or an explicit proxy pipeline. Keep the boundary visible in logs and UI so users know whether DeepSeek processed the original image or only a generated text description.

For teams tracking the research, the next meaningful launch signals are an official model card or API model ID, a documented image request schema, public weights, image-token pricing, and an update to the supported-feature table. Until one of those appears, DeepSeek Vision is a research direction rather than a public V4 endpoint.