cd /news/artificial-intelligence/deepseek-s-experimental-vision-model… · home topics artificial-intelligence article
[ARTICLE · art-105812] src=runtimewire.com ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

DeepSeek's experimental vision model spans three formats, caps images at 384 tokens

DeepSeek announced deepseek-v4-flash-vision-exp on August 21, 2026, an experimental vision model supporting Chat Completions, Messages, and Responses request formats, with each image billed at no more than 384 V4-Flash tokens. The production V4-Flash endpoint remains text-only, and DeepSeek has published no benchmark, latency, or reliability data for the vision model.

read3 min views5 publishedAug 21, 2026
DeepSeek's experimental vision model spans three formats, caps images at 384 tokens
Image: Runtimewire (auto-discovered)

Liang Wenfeng's separately named deepseek-v4-flash-vision-exp supports Chat Completions, Messages and Responses requests, with each image billed at no more than 384 V4-Flash tokens.

By RuntimeWire Staff · Published

Primary source: DeepSeek

Why it matters #

DeepSeek has attached a predictable image-input ceiling to a separately named experimental model available through three familiar request formats. The production V4-Flash endpoint remains text-only, and DeepSeek has published no benchmark, latency or reliability data for evaluating the vision model.

Liang Wenfeng's DeepSeek announced deepseek-v4-flash-vision-exp on August 21, 2026, an experimental model for mixed text and image requests. The announcement names three compatible request formats - Chat Completions, Messages and Responses - and caps each image at 384 billable tokens under V4-Flash pricing.

The release concerns the request contract and billing of this separately named vision model. It uses a separate experimental model rather than changing the production V4-Flash endpoint, according to DeepSeek's vision guide.

Liang built DeepSeek after applying machine learning to a different numbers business. He founded the AI-driven investment firm Jacobi in 2013 and co-founded quantitative hedge fund High-Flyer two years later, according to TechCrunch's account of his early career. Liang earned bachelor's and master's degrees in engineering from Zhejiang University, according to Forbes. He launched DeepSeek in 2023 after High-Flyer had supplied the capital and computing base for its research.

A 384-token billing ceiling

DeepSeek's current pricing table lists V4-Flash at $0.22 per million uncached input tokens and $0.66 per million output tokens during off-peak hours. Cached input costs $0.007 per million tokens off-peak. Peak rates are $0.44 per million uncached input tokens, $1.32 per million output tokens and $0.014 per million cached input tokens. DeepSeek says the experimental vision model uses those same rates.

At the off-peak uncached input rate, the maximum 384-token image charge would be $0.00008448. The same image-token ceiling would cost $0.00016896 at the peak rate. Those calculations cover one image input; accompanying text, generated output and additional calls add to the bill. Image dimensions determine the actual token conversion, so 384 tokens functions as a ceiling instead of a flat per-image fee.

The cap gives developers an upper bound for the image portion of a request. It provides no evidence about how much visual detail survives tokenization, especially in screenshots with small text, diagrams or dense interface elements.

One model, three request formats

DeepSeek is distributing deepseek-v4-flash-vision-exp through request shapes already used by developers. Its announcement names OpenAI-compatible Chat Completions, Anthropic-compatible Messages and Responses. Images can be included as base64 data or URLs alongside text.

The three-format support applies specifically to deepseek-v4-flash-vision-exp . Developers must select that model; the production V4-Flash endpoint does not gain vision support from this announcement.

The interface choices map to established API conventions. OpenAI's Responses API accepts image inputs alongside text and files. Anthropic's vision documentation supports base64 images and hosted URLs. Mistral's vision interface takes URLs or base64 data through Chat Completions. DeepSeek's specific proposition is one experimental model exposed through all three request formats with a stated image-token ceiling.

The launch announcement does not include image-accuracy benchmarks, latency measurements, reliability figures, supported resolutions, file-size limits or a maximum number of images per request. DeepSeek also has not said in the cited materials whether it will release weights for the experimental vision model.

Vision branches off from V4-Flash

DeepSeek introduced the broader V4 family on April 24 with a 1 million-token context window and open weights for V4-Pro and V4-Flash, according to its V4 Preview announcement. The August release attaches vision to a separate experimental endpoint without changing the production V4-Flash model.

In June 2026, DeepSeek raised more than $7.4 billion at a valuation above $50 billion, according to Forbes and The Information. Tencent, CATL, JD.com and NetEase were among the reported investors.

For developers, the test is narrowly defined: whether a model capped at 384 image tokens can read the screenshots, documents and interface states their applications encounter. DeepSeek has disclosed the three request formats and billing ceiling while leaving the performance case for users to measure.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @deepseek 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/deepseek-s-experimen…] indexed:0 read:3min 2026-08-21 ·