DeepSeek V4 Flash Can Finally See — Here's What Changed
Every model in DeepSeek's V4 Flash line has been text-only since it launched. That changed on August 21, 2026, when DeepSeek quietly shipped DeepSeek-V4-Flash-Vision-Exp, an experimental build that reads images, screenshots, and charts, at the same rock-bottom price as the text-only version. Here's what it actually adds, how it stacks up against Claude Opus 4.8 on DeepSeek's own numbers, and what \"experimental\" means for anyone thinking about building on it.
DeepSeek-V4-Flash-Vision-Exp adds native image input to the V4 Flash line for the first time.
V4 Flash Can Now Read What You Show It
DeepSeek-V4-Flash-Vision-Exp is exactly what the name suggests: an experimental version of DeepSeek's V4 Flash model with image understanding added on top. You can now send it screenshots, charts, diagrams, and photos alongside text, and it reasons about them directly, rather than routing images through a separate captioning step first. DeepSeek says the model matches standard V4 Flash on every text-only benchmark, reasoning, agent tasks, world knowledge, so you're not trading away text quality to get vision.
It ships through DeepSeek's API under the model ID deepseek-v4-flash-vision-exp, works across the Chat Completions, Messages, and Responses formats, and accepts images as inline base64 data, external URLs, or references to DeepSeek's newly opened, free-to-use Files API, which lets you upload an image once and reuse it across multiple requests without re-uploading.
How Close Is It to Claude Opus 4.8?
A real jump over the text-only model
On DeepSeek's published multimodal agent benchmarks, Vision Exp shows a substantial leap over standard V4 Flash, which can't process images at all and simply ignores any visual elements in these same tests.
Trading wins with Opus 4.8, not sweeping it
DeepSeek leads on Agents' Last Exam and ZeroBench; Opus 4.8 stays ahead on ApexBench and Chartography. "Close to Opus 4.8" is a fair read of these specific results, not proof of a clean win across every visual task.
Worth remembering: Opus 4.8 isn't Anthropic's newest model
Anthropic released Claude Opus 5 on July 24, 2026. The comparison DeepSeek is making is against Anthropic's previous flagship, not its current one.
These are provider-published numbers
The benchmarks above come from DeepSeek's own launch materials. Independent testing on real-world OCR, small-text reading, visual hallucination rates, and long agent runs is still the gap to watch before trusting these figures blindly.
DeepSeek's own launch demos included redesigning a website from a screenshot and building a presentation from visual input.
The Pricing Is the Real Headline
Vision at text-model prices
$0.14 per million cache-miss input tokens and $0.28 per million output tokens, identical to text-only V4 Flash. Most vision-capable models charge a real premium over their text-only counterparts.
A hard cap on image cost
Each image is billed at up to 384 tokens, regardless of resolution, which makes cost predictable for agents that repeatedly look at screenshots or document pages.
Free Files API
Upload an image once and reference it by file ID across multiple requests, cutting repeated upload bandwidth for workflows that reuse the same reference images.
Real use cases, not just a demo
DeepSeek's own launch examples include building a commercial travel presentation, redesigning a website from a screenshot, and generating an animated front-end demo from a visual mockup.
"Experimental" and "API-Only" Both Matter Here
DeepSeek is explicit that this is an experimental build, not a production-committed release, and it isn't billed as a replacement for standard V4 Flash. As of this writing, it's API-only, with no downloadable weights on Hugging Face, unlike much of DeepSeek's other V4-generation work. Community expectation is that an eventual production vision model may go open-weight, but that's unconfirmed, not a promise DeepSeek has made.
Bottom Line
This is DeepSeek closing a real gap, not a marketing stunt: V4 Flash going from text-only to a genuinely capable vision model, at the same price, is a meaningful jump for anyone building cost-sensitive agents that need to read screenshots or documents. The benchmark comparisons to Opus 4.8 are directionally fair but shouldn't be read as a clean victory, and the experimental, API-only status means this isn't the moment to bet a production system on the current model ID. Worth testing against your own images now; worth waiting for the production release before relying on it.
More AI Guides, Written Clearly
Browse our full library of AI tool guides, updated regularly as models and features change.
Common Questions
What is DeepSeek-V4-Flash-Vision-Exp?
It's an experimental multimodal version of DeepSeek's V4 Flash model, released August 21, 2026, that adds native image understanding, screenshots, charts, diagrams, and photos, while matching standard V4 Flash on text-only tasks.
How much does DeepSeek V4 Flash Vision cost?
The same as standard V4 Flash: $0.14 per million cache-miss input tokens and $0.28 per million output tokens. Each image is billed at up to 384 tokens regardless of resolution.
Is DeepSeek V4 Flash Vision as good as Claude Opus 4.8?
DeepSeek's own benchmarks show it trading wins with Opus 4.8, leading on some multimodal agent tests and trailing on others, so it's fair to call it close, not clearly better. Note that Opus 4.8 is Anthropic's previous flagship, not its current one, Claude Opus 5.
Can I download DeepSeek V4 Flash Vision's weights?
Not yet. As of its August 21, 2026 launch, it's API-only with no open weights published, unlike some of DeepSeek's other V4-generation releases.
More from the creator
Explore two other projects from the same creator.