DeepSeek · 2026-08-21 · major
DeepSeek V4-Flash-Vision-Exp — an experimental V4 model that reads images
DeepSeek V4-Flash-Vision-Exp is an experimental multimodal model on the DeepSeek API that takes images alongside text. It scores 83.9 on Terminal Bench 2.1 and matches DeepSeek V4-Flash on text-only work.
An experimental DeepSeek model that adds image understanding to the V4-Flash line, at V4-Flash prices.
Quick facts
| Maker | DeepSeek |
|---|---|
| Model ID | deepseek-v4-flash-vision-exp |
| Input types | Text + images (JPEG, PNG, GIF, WebP) |
| Context window | 1M tokens |
| Status | Experimental |
| Image billing | Up to 384 tokens per image |
| APIs | Chat Completions, Messages, Responses |
Pricing
| Input (cache miss) · Off-peak; peak is double | $0.22 / 1M tokens |
|---|---|
| Input (cache hit) · Off-peak; peak is double | $0.007 / 1M tokens |
| Output · Off-peak; peak is double | $0.66 / 1M tokens |
What is it?
DeepSeek V4-Flash-Vision-Exp adds image input to DeepSeek's V4-Flash line. The model takes mixed text-and-image prompts through the Chat Completions, Messages and Responses APIs, so an agent can read a screenshot, a chart or a photo in the same call it reads text. DeepSeek labels the model experimental and ships it under the id deepseek-v4-flash-vision-exp.
How does it work?
Images are turned into tokens and billed like text, and can be passed as base64 data, as an external URL, or through the DeepSeek Files API. On pure-text work — agents, reasoning, world knowledge — DeepSeek V4-Flash-Vision-Exp performs on par with the official DeepSeek V4-Flash, so sight is added without a text-quality trade-off. DeepSeek Harness 0.1.1 shipped the same day with out-of-the-box support for the new model.
Why does it matter?
Agent benchmarks that need visual understanding are where DeepSeek V4-Flash-Vision-Exp gains the most. DeepSeek says the jump over DeepSeek V4-Flash is significant and brings its multimodal agent skills close to Opus-4.8, at flash-tier prices. Teams building screen-reading, chart-parsing or document agents get a much cheaper model to test against.
Who is it for?
developers building vision-capable AI agents
Frequently asked questions
- How much does DeepSeek V4-Flash-Vision-Exp cost?
- DeepSeek V4-Flash-Vision-Exp is billed at DeepSeek V4-Flash rates. Off-peak, input is $0.22 per million tokens on a cache miss and $0.007 on a cache hit, and output is $0.66 per million tokens. Peak rates are double the off-peak rates, and DeepSeek sets peak hours at 01:00-04:00 and 06:00-10:00 UTC.
- Is DeepSeek V4-Flash-Vision-Exp stable enough for production?
- DeepSeek labels V4-Flash-Vision-Exp an experimental model and gives it an -exp suffix in its id, so treat it as a preview rather than a settled endpoint. Because its text performance matches the official DeepSeek V4-Flash, a team can trial vision calls on the experimental id while keeping text traffic on the stable V4-Flash model.
- How does DeepSeek V4-Flash-Vision-Exp compare to DeepSeek V4-Flash?
- On pure-text work — agents, reasoning and world knowledge — DeepSeek V4-Flash-Vision-Exp performs on par with the official DeepSeek V4-Flash. The gap shows on agent benchmarks that need visual understanding, where DeepSeek says V4-Flash-Vision-Exp makes a significant leap over V4-Flash and comes close to Opus-4.8. Price and the 1M-token context window are the same.
- Which benchmark scores did DeepSeek publish for V4-Flash-Vision-Exp?
- DeepSeek listed nine results for V4-Flash-Vision-Exp in its change log: Terminal Bench 2.1 at 83.9, NL2Repo at 57.7, DeepSWE at 59.3, DSBench-Hard at 63.6, AutomationBench (public) at 25.7, ApexBench pass@1 at 36.5, Agents' Last Exam at 27.3, Chartography at 64.3, and ZeroBench pass@5 at 35.0.
Try it
model='deepseek-v4-flash-vision-exp'