DeepSeek-V4-Flash-Vision-Exp goes live with image input for V4-Flash agents (2026)

DeepSeek made DeepSeek-V4-Flash-Vision-Exp live on its API Platform on August 21, adding image input to the V4-Flash line. The experimental model keeps DeepSeek's claimed V4-Flash text capabilities for agents, reasoning, and world knowledge. DeepSeek says its multimodal agent results move close to Opus-4.8; that comparison comes from DeepSeek's own evaluation table. 1 2

What launched

Signal	Confirmed detail	Why it matters
Release	The API model ID is `deepseek-v4-flash-vision-exp`; DeepSeek Harness 0.1.1 adds out-of-the-box support. 1	Developers can start with an existing agent harness instead of waiting for a separate integration.
Inputs	The model accepts mixed text and images through Chat Completions, Messages, and Responses. Images can arrive as base64 data, public URLs, or Files API references. 3	Screenshot, chart, document, and visual tool workflows can share one endpoint.
Cost boundary	DeepSeek says images are billed at V4-Flash pricing and capped at 384 tokens per image after resizing. The Files API supports reusable uploads up to 64 MiB. 1 4	Image count and file size still affect request design and spend.
Action window	The release materials document hosted API access and a Files API. A local-weight deployment path is absent from this launch. 1	API users can test now; teams planning self-hosting need more information.

What the benchmark table says

DeepSeek's comparison of V4-Flash-Vision-Exp, V4-Flash 0731, and Opus-4.8 across text-based and multimodal agent evaluations — DeepSeek's published comparison reports the scores and evaluation conditions; the footnote says its code-agent tests used DeepSeek Harness Minimal Mode, maximum output tokens, `top_p=0.95`, and `temperature=1.0`. 1

The multimodal scores are the launch's main claim. Vision-Exp scores 36.5 on ApexBench versus 26.2 for V4-Flash 0731 and 27.3 on Agents' Last Exam versus 25.2. The model also reaches 64.3 on Chartography and 35.0 on ZeroBench, close to Opus-4.8's 65.0 and 34.0. The same table shows a mixed text-agent result: Vision-Exp leads the older Flash model on Terminal Bench 2.1 and DeepSWE, while trailing it on Cybergym. 1

Why this matters

DeepSeek has added a hosted multimodal option to a model line that already targets reasoning and tool use. OpenRouter's independent listing reports a 1,048,576-token context window and one provider on its platform, adding a second hosted access route for developers. 5

A Hacker News thread linking to DeepSeek's Vision guide had 388 points and 126 comments when retrieved, an early attention signal rather than a quality verdict. 6 The practical next step is a small test loop with screenshots, charts, or visual tool outputs. Treat DeepSeek's benchmark table as a set of hypotheses to check against the workflows your team actually runs.