DeepSeek V4 Flash Vision Exp

DeepSeek launched a new model on its API platform on August 21: DeepSeek-V4-Flash-Vision-Exp. The name is a mouthful, but the idea is simple: V4-Flash can now see images.
Advertisement 728 × 90
CompanyDeepSeek
Context1M
Released2026-08
Updated2026-08-28

DeepSeek V4 Flash Vision Exp Overview

This is not a completely separate vision model.

It is still based on V4-Flash, with image understanding added on top of the existing text, reasoning, Agent, and knowledge capabilities.

A practical way to think about it is:

V4-Flash + image understanding.

The “Exp” at the end stands for Experimental, so this is still an experimental release.

Under the hood, it uses a sparse MoE architecture with 284B total parameters and 13B active parameters per request.

Those numbers matter less than the practical takeaway: DeepSeek added vision without turning it into a much more expensive model to run.

It supports JPEG, PNG, GIF, and WebP images, as well as mixed text-and-image prompts.

The context window goes up to 1M tokens, with a maximum output of 384K tokens.

On the API side, it still works with features such as:

JSON Output
Tool Calls
Responses API
Anthropic-compatible API formats

So this is not just a new “image upload” option in a web interface.

DeepSeek has brought vision directly into the API, which makes it much more useful for developers building Agents and automated workflows.

DeepSeek V4 Flash Vision Exp Pricing

PlanPriceDescription
Cached input RMB 0.05 / 1M tokens off-peak; RMB 0.10 / 1M tokens peak Best for repeated context, fixed prompts, or reused input
Uncached input RMB 1.50 / 1M tokens off-peak; RMB 3.00 / 1M tokens peak New input is billed here; image tokens are also counted as input
Output RMB 4.50 / 1M tokens off-peak; RMB 9.00 / 1M tokens peak Charged based on the number of tokens the model generates
Files API Free Upload an image once and reuse it with a file_id; up to 64 MiB per image

Images are not billed per file.

They are converted into tokens, and a single image can use up to 384 tokens.

At the peak uncached-input rate:

384 × RMB 3 ÷ 1,000,000 ≈ RMB 0.00115

So even at the maximum image token count, the input cost for one image is only around one-thousandth of a yuan.

That is extremely cheap.

The concurrency limit remains 2,500, the same as V4-Flash.

The Files API is also free. Once an image is uploaded, later requests can reference it with a file_id instead of sending the same file again.

If an Agent needs to inspect the same screenshot, chart, or design multiple times, this is much cleaner than repeatedly sending Base64 data.

DeepSeek V4 Flash Vision Exp Key Features

Three ways to send images

There are three supported methods, and none of them are unusual for developers.

Base64 inline

You can encode a local image and include it directly in the request.

It is easy for testing and small files, but the full request body has a 48 MiB limit. Base64 also makes files larger, so it is not the best choice for repeated use.

External URL

You can pass a publicly accessible image URL and let the service fetch it.

This is convenient, but the image must be reachable from the internet and cannot exceed 32 MiB.

Files API

Upload the image first, then refer to it later using a file_id.

This supports files up to 64 MiB and is the most practical option when the same image will be used more than once.

Agents can finally see their own output

This is the part that matters most.

Basic image understanding is not new anymore. The bigger change is that vision can now be used inside Agent workflows.

DeepSeek showed examples such as:

  • generating a premium Tibet road-trip presentation
  • redesigning the DeepSeek Harness website in a different visual style
  • building a clay-monster-style front-end demo

These are not simple “describe this image” tasks.

They require the model to work across images, code, tools, and final visual output.

That matters because text-only Agents have a major weakness: they can build something without being able to see how it actually looks.

They can write code.

They can edit code.

They can call tools.

But if a web page looks terrible, a chart is unreadable, or a slide deck has broken spacing, the Agent cannot judge that on its own.

That is a real limitation.

V4-Flash-Vision-Exp fixes part of that problem.

The Agent can now look at what it created, then make another pass.

For web pages, presentations, front-end work, charts, and UI tasks, that is much more useful than basic image captioning.

Benchmark scores also improved

DeepSeek published several multimodal Agent benchmark results.

ApexBench (Pass@1)

  • V4-Flash-Vision-Exp: 36.5
  • V4-Flash: 26.2
  • Claude Opus-4.8: 39.4

Agents' Last Exam

  • V4-Flash-Vision-Exp: 27.3
  • V4-Flash: 25.2
  • Claude Opus-4.8: 25.7

Chartography

  • V4-Flash-Vision-Exp: 64.3
  • Claude Opus-4.8: 65.0

ZeroBench (Pass@5)

  • V4-Flash-Vision-Exp: 35.0
  • Claude Opus-4.8: 34.0

Across these four tests, DeepSeek and Claude Opus-4.8 trade wins rather than showing a clear overall winner.

One detail is worth keeping in mind.

ApexBench and Agents' Last Exam include multimodal content, while the older V4-Flash model cannot process images and simply ignores that information.

So the jump from 26.2 to 36.5 on ApexBench does not mean the model suddenly became dramatically smarter.

A big part of the gain comes from the fact that it can finally access information that was previously invisible to it.

Some text-only Agent benchmarks also improved:

  • Terminal Bench 2.1: 82.7 → 83.9
  • DeepSWE: 54.4 → 59.3
  • DSBench-Hard: 59.6 → 63.6

At least from these numbers, adding vision did not hurt the model’s existing Agent performance.

It fixes one of V4’s biggest gaps

V4-Flash-0731 was a text-only model.

That is fine for chat, but the limitation becomes obvious once you use it as an Agent.

It cannot inspect a screenshot.

It cannot read a chart visually.

It cannot check whether a UI layout broke.

It cannot look at a finished slide deck and judge whether the design actually works.

That is the bigger story here.

A lot of automated workflows hit the same wall: the Agent can create something, but it cannot see the result.

With vision added, it can at least inspect the output and decide what needs another pass.

For web, UI, slide, and chart-heavy workflows, that is a meaningful upgrade.

Summary

If you already use V4-Flash for Agent development, this version is worth testing.

The API does not change much, vision is added, and the pricing stays the same.

That combination is especially attractive for tasks involving screenshots, websites, dashboards, presentations, charts, and UI.

There is still one obvious caveat: the model has Exp in the name for a reason.

It is an experimental release, so reliability and edge-case performance still need more real-world testing.

Face and identity recognition is one area to be careful with. There have been tests where the model misidentified Liang Wenfeng, so this is not a good fit for tasks that require highly reliable person identification.

It is also API-only, with no separate web interface.

If you only want AI to read the occasional image, there is little reason to set up an API workflow just for this.

But if you are already building Agents that work with screenshots, websites, charts, or visual interfaces, V4-Flash-Vision-Exp is much more interesting.

V4-Flash basically got a pair of eyes without making the bill any bigger.

Comments (0)

Leave a comment

Advertisement 728 × 90

DeepSeek Model Comparison

Model Context Pricing API Released Global Heat
DeepSeek V4 Flash Vision Exp
1M YES 2026-08
1M YES 2026-08
96/100
1M YES 2026-07
87/100
1M NO 2026-04
90/100
128K Freemium YES 2025-05
96/100

Similar Models

DeepSeek R2
96
Open-source specialized reasoning model, with MATH benchmark test scores comparable to GPT-4o, released under the MIT open-source license.
DeepSeek
DeepSeek V4 Pro 0813
96
DeepSeek V4 Pro 0813 feels more like a workhorse model to me. For casual chat or light editing, I wouldn’t bother. Its value shows up when I’m digging through code, chasing bugs, or handling tightly connected tasks.
DeepSeek
DeepSeek V4 Pro(Legacy)
90
DeepSeek’s flagship MoE model designed for complex reasoning, advanced coding, AI Agents, and long-context professional workflows.
DeepSeek
DeepSeek V4 Flash 0731
87
DeepSeek’s advanced AI model optimized for coding, AI Agents, tool calling, complex reasoning, and large-scale AI applications with million-token context support.
DeepSeek
Claude Opus 5
100
Anthropic’s flagship Claude model built for complex reasoning, AI Agents, software development, and enterprise knowledge workflows.
Anthropic
GPT-5.5 Pro
99
I’ve been using GPT-5.5 Pro with a few colleagues for the past several weeks, mostly on the kinds of jobs where regular chatbots tend to fall apart: long documents, messy research questions, code debugging, and tasks that need more than one or two steps of reasoning.
OpenAI
GPT-5.5
99
OpenAI flagship model, a multimodal AI natively supporting text, images, and audio.
OpenAI
GPT-5.6 Luna
99
An efficient GPT-5.6 series model optimized for speed, cost efficiency, and intelligent performance across large-scale AI applications.
OpenAI

Related News