On August 21, DeepSeek officially launched the V4-Flash-Vision-Exp API service for its multimodal vision understanding model. The model is priced the same as V4-Flash, consuming 384 tokens per image—less than half of competitors like GPT and Claude. The official team stated that its multimodal agent capabilities are nearing Opus-4.8 levels. Real-world tests show the model can generate executable code from screenshots within 30 seconds, covering scenarios such as web pages and games. Notably, the thinking mode consumes the entire output budget; developers are advised to disable it to ensure proper output. DeepSeek chose to pilot this model on the more cost-effective Flash version rather than the Pro version, aiming to equip agents within its ecosystem with genuine visual understanding capabilities.

Author and source: GeekPark

Multimodal is also at Liang Xin's bottom price.

The whales have awakened!

Suddenly, on August 21, DeepSeek officially announced the launch of V4-Flash-Vision-Exp and the availability of its multimodal API service.

After waiting so long, the DeepSeek model has finally gained multimodal capabilities, and excited users have almost forgotten that DeepSeek just started increasing prices based on usage time.

Liang Wenheng adjusted the potentiometer by turning the knob one notch toward the "Liang Sheng" direction.

GeekPark couldn't wait and immediately ran a test.

Columnist recruitment

Geek Park’s new column, “AI New Arrivals,” will introduce you to the latest AI applications and hardware, helping you become the coolest person in the AI era!

We now invite all students who enjoy trying out new technologies and experiencing AI to contribute. If you discover and test a new AI application or feature, submit your review in the following format (see example: “Tested DeepSeek V4 Pro Official Release: Enhanced Agent Capabilities | New AI Feature”) to our column for publication on the GeekPark WeChat official account. Not only will you receive a writing fee, but we’ll also reimburse you for the subscription cost of the AI application.

Meanwhile, outstanding authors will have the opportunity to join the GeekPark AI Experience Group, gain early access to the latest AI applications and tools, attend exclusive GeekPark AI events, and engage in one-on-one conversations with founders of AI applications.

AGI is taking too long—let’s seize the day and get some people started with AI now! To submit content or join the group, scan the QR code below to add the Geek Assistant on WeChat👇

01

Performance close to Opus 4.8

DeepSeek has announced on its WeChat official account that the new multimodal vision understanding model, DeepSeek-V4-Flash-Vision-Exp, is now live on the API platform. Developers can directly invoke it by setting model='deepseek-v4-flash-vision-exp', with support for three image input methods: base64 inline, external URLs, and Files API.

But the presence of "Exp" in the name suggests the story has only just begun.

According to the official announcement, several key parameters of V4-Flash-Vision-Exp are worth noting.

Data provided by DeepSeek | Image source: DeepSeek

Priced the same as V4-Flash. Each image consumes up to 384 tokens, billed per token, with no additional visual processing surcharge. For comparison, GPT and Claude typically consume 800 to 1,100 tokens for images of the same resolution. For the same image, DeepSeek’s token cost is less than half that of its competitors. This efficiency stems from DeepSeek’s previously disclosed “Thinking with Visual Primitives” framework, which compresses visual elements into spatial-coordinate-based primitives rather than splitting the entire image into dense patches.

Text-only capabilities have not been degraded. The official statement emphasizes that Vision-Exp matches the official V4-Flash version on text-only benchmarks such as Agent, reasoning, and world knowledge. This means it is not a specialized model that sacrifices text performance for image understanding, but rather a model that adds visual comprehension on top of the Flash foundation.

The multimodal agent capability has "nearly reached Opus-4.8." This is the boldest statement in the official announcement. On agent benchmarks requiring visual understanding, Vision-Exp achieves a significant leap compared to the text-only V4-Flash.

Also launching同步 is the Files API—this interface is free to use. Developers can first upload images to the platform to obtain a file_id, then directly reference it in subsequent requests, eliminating the need to retransmit the same image multiple times. This is a practical optimization for Agent workflows that require repeated analysis of the same set of images.

02

Create an AI remake of "Niu Lai"

Now that we've covered the specs, let's put it to the test.

What was the biggest meme of the 2026 summer vacation? "Niu Lai." This animated film, painstakingly handmade by a mother and son over five years with crude 3D modeling that netizens called "even AI couldn't generate," saw its box office surge from just 7,169 yuan to 30 million, becoming the internet's largest piece of performance art.

We decided to use it for our first test—feeding DeepSeek Vision a classic close-up screenshot of “Niu Lai” saying “Mom,” and asking it to understand the image and recreate it using pure code.

Comparison image_NiuLai_Restored.png — Left: Original screenshot, Right: AI-restored version | Source: GeekPark

DeepSeek's performance is somewhat hard to describe.

But at least it accurately identified the bull’s orange color scheme, furrowed brows, thick lips, and its “fierce yet pitiful” expression, while preserving the subtitle “Mom” at the bottom of the image. The rendered SVG + CSS code produces a cartoon bull with gradient highlights and smooth, rounded lines—jumping directly from the “uncanny valley” to “kawaii.”

The entire process took only 35 seconds, and the image consumed just 384 tokens. Netizens said, “AI looks more like a movie than the original,” and now there’s proof (just kidding).

We also conducted a more interesting test—showing DeepSeek a screenshot of two cows facing each other in "Niu Lai," and asking it to directly create an interactive endless runner game based on the cow imagery in the image.

Cow Parkour Game | Image Source: GeekPark

DeepSeek extracted the color scheme and basic shape of the bull from the screenshot and used CSS to create a pixel-style running bull character, along with a sky, grass, obstacles, and a scoring system. From seeing the image to outputting a fully playable HTML game, it took just 36 seconds.

This demo itself may not have direct productivity value, but it demonstrates something more important—when visual understanding and code generation capabilities are combined, AI can achieve far more than with either capability alone.

03

Write the code based on the screenshot, directly and completely.

The second test better reflects real-world work scenarios.

We created a landing page design for a payment product—featuring a dark gradient background, navigation bar, large headline, syntax highlighting in a code block, dual buttons, and three feature cards—then took a screenshot and fed it to DeepSeek Vision, requesting it to directly output runnable HTML.

Comparison image: Left side is the design mockup, right side is the HTML page generated by AI | Image source: GeekPark

After 26 seconds, DeepSeek generated a complete responsive HTML page. The gradient background, navigation bar layout, button styles, and syntax highlighting colors for the code area—all matched precisely. It even autonomously added hover effects and entrance animations.

This capability directly addresses the core use cases of an AI Agent. When your AI Agent needs to interact with a browser, understand the structure of a web interface, or quickly convert a designer’s mockup into frontend code, visual understanding is no longer a "nice-to-have"—it’s a prerequisite for getting the job done.

However, a familiar issue was encountered during testing.

When the default thinking mode is enabled, the model’s reasoning tokens balloon uncontrollably, consuming the entire output budget. In our scenario understanding tests, we observed that all 2,001 completion tokens were reasoning tokens—resulting in zero actual output. The model spent considerable time thinking but produced not a single word.

The solution is the same as last time—set reasoning_effort: "none" to disable the reasoning mode. After disabling it, the same test time dropped from 23 seconds to 7.9 seconds, with complete and high-quality output.

This is the same issue reproduced across two consecutive versions. When the V4 Pro official release was launched, we found that the thinking mode consumed over 82% of the output budget; now, Flash Vision is experiencing 100% of the budget being consumed by reasoning. For developers, this means a practical, hard rule: when performing vision tasks, always disable the thinking mode or set a sufficiently large max_tokens value, or you will likely receive an empty response.

We've already walked this path twice for you.

04

Why Flash?

A thoughtful product decision is that DeepSeek chose to initially test its visual capabilities on Flash rather than the flagship Pro.

From a cost perspective, Flash is a 284B-parameter MoE model with significantly lower inference costs compared to the 1.6T-parameter Pro. Each image consumes at least 384 tokens of the input budget; when layered on top of Pro’s pricing, costs escalate rapidly. Flash’s low-cost foundation makes multimodal usage economically viable.

From a timing perspective, just the day before, the DeepSeek Harness framework released version RC.8, adding native support for image requests.

Harness has connected the pipeline for multimodal agents at the framework level, while Flash Vision has filled in the visual understanding engine at the model level—both moves are aimed at the same goal:

Enable agents in the DeepSeek ecosystem to truly “see.”

Looking back at the four-month multimodal learning timeline—late April introduced grayscale image recognition, May saw broad rollout on the web, June brought full launch on the app, and August added multimodal support to the Harness framework—the same day Flash Vision API was opened. Each step built precisely on the previous one. Yet, instead of a launch event, it quietly opened the door with a model name ending in “Exp.”

For developers who have been waiting for a multimodal API, the wait is over—even though it’s labeled “Exp” rather than “GA.” For teams building multimodal agents, a vision model that consumes less than half the tokens of competitors while maintaining unaltered text capabilities may be the most cost-effective option available today.

Just don't forget to turn off the thinking mode.

*Header image source: AI-generated

This article is originally published by GeekPark. Please contact GeekPark's WeChat: geekparkGO.