何が起こったのか
KuCoin reports, citing GeekPark, that DeepSeek launched the experimental V4-Flash-Vision-Exp API on August 21. The reported service accepts images through inline base64 data, external URLs and a Files API. GeekPark’s tests described image-to-code tasks, but the source does not provide independent verification of DeepSeek’s stated benchmarks, pricing or general availability.
KuCoin’s August 24 page reports that DeepSeek made V4-Flash-Vision-Exp available through an API on August 21, citing the technology publication GeekPark and an announcement attributed to DeepSeek’s official WeChat account. The model name includes “Exp,” indicating an experimental release rather than a clearly documented general-availability launch. The report says developers can invoke it with the model identifier “deepseek-v4-flash-vision-exp.” The source does not independently link to a public DeepSeek API specification or establish how broadly the service was available at the time of publication.
The reported interface accepts three forms of image input: inline base64 data, external URLs and DeepSeek’s Files API. KuCoin says the model is priced on the same basis as V4-Flash and uses 384 tokens per image, without a separate visual-processing surcharge. GeekPark compares that figure with reported consumption of 800 to 1,100 tokens for images of similar resolution in GPT and Claude systems, and says DeepSeek’s image cost is less than half. Those comparisons are claims reproduced by the source, not results independently measured in this review. The report also says the Files API lets developers upload an image once, receive a file identifier and reuse it in later requests.
GeekPark’s reported demonstrations focused on converting images into executable code. In one test, the model analyzed a screenshot from the animated film “Niu Lai” and generated SVG and CSS intended to reproduce the pictured bull; the report says the process took 35 seconds. In a second demonstration, it used a screenshot of two cows to create a simple endless-runner game in HTML and CSS, with the report giving a 36-second completion time. A third test used a payment-product landing-page screenshot containing a dark , navigation, a headline, a code block, buttons and feature cards. GeekPark says the model generated responsive HTML in 26 seconds and reproduced several visual elements. These are demonstrations by GeekPark, not controlled benchmarks, and the source does not establish their hardware, network conditions, evaluation criteria or repeatability.
The report identifies a significant operational limitation. According to GeekPark, enabling the default thinking mode caused all 2,001 completion tokens in one test to be consumed by reasoning, leaving no visible answer. The publication says that disabling reasoning with “reasoning_effort: none” reduced one test from 23 seconds to 7.9 seconds and produced a complete response. KuCoin also repeats GeekPark’s claim that the model’s multimodal agent performance was “nearly” at the level of Opus-4.8 and that text-only performance remained comparable with V4-Flash. Neither claim is independently confirmed in the source.
なぜそれが重要なのか
If the reported token efficiency and visual performance hold up, the service could make image-aware AI agents less expensive to operate, particularly for browser, interface-understanding and code-generation tasks. The practical value remains uncertain because the report is based on a single publication’s tests and does not establish reliability across broader workloads.
The central significance is economic as well as technical. Image-aware agents must repeatedly interpret screenshots, documents or visual interfaces before they can act. If each image consumes materially fewer input tokens, the cost of tasks such as browser navigation, interface testing and screenshot-to-code conversion could fall. That could make visual capabilities practical in more agent workflows, especially where an agent examines many images or revisits the same visual context through a file reference.
The reported Files API could also reduce unnecessary retransmission. Reusing an uploaded image through a file identifier may simplify workflows that require repeated analysis, although the source does not document retention periods, access controls, deletion procedures, file-size limits or whether uploaded images are used for training. Those omissions matter for organizations handling private designs, customer information or other sensitive visual data. The report provides no independent privacy or security assessment.
The launch is also relevant to competition among lower-cost multimodal models. DeepSeek is reported to have placed the vision capability on its Flash model rather than its larger Pro model, a choice the article frames as a way to keep costs manageable. But the comparison with GPT, Claude and Opus-4.8 is not a like-for-like evaluation in the supplied material. Token accounting can vary by image size, detail level, encoding and provider, while visual quality depends on the task. The report therefore supports interest in the product’s direction, not a conclusion that it is already cheaper or better for general use.
インタラクティブなメカニズム: 実際にどのように機能するか
この開発の背後にある基盤となるテクノロジーをインタラクティブに探索します。
Which component of an AI application is the machine-learning model itself?
次に見るべきもの
Developers and users should look for primary documentation confirming access, regional availability, pricing, limits, supported image formats and Files API behavior. Independent evaluations should test visual reasoning, generated code, latency, failure rates and performance with the reasoning mode enabled and disabled. The main unresolved question is whether the reported cost and capability claims generalize beyond the demonstrations described by GeekPark.
The first verification priority is primary product documentation. Users need confirmation of the API’s public status, supported regions, pricing formula, image-resolution limits, rate limits, maximum output length, Files API retention and deletion behavior, and whether the experimental service can be used in production. The source describes the service as live but does not establish a stable service-level commitment or a general-availability date.
Independent testing should separate the model’s visual understanding from its code-generation ability. Useful evaluations would repeat the reported screenshot-to-HTML tasks across different layouts, image qualities and interfaces, then measure functional correctness rather than visual resemblance alone. Tests should also compare latency, token use and error rates with and without thinking mode. The report’s empty-output example makes configuration an immediate practical concern, but it does not show how often that failure occurs across requests.
Reliability and safety remain open questions. A model that can read a screenshot and generate executable code may help with accessibility, interface migration and rapid prototyping, but generated code can contain functional, security or accessibility defects. The supplied report gives no evidence about these risks, no systematic red-team results and no information about handling images containing personal or confidential data. Until those questions are answered, the most defensible view is that DeepSeek has reported an experimental multimodal API with promising demonstrations and important limits, rather than a verified replacement for established vision systems.