DeepSeek-VL2

Overview

DeepSeek-VL2 is the successor to DeepSeek-VL, DeepSeek's earlier vision-language series. Its language side is a DeepSeekMoE mixture-of-experts model, and images pass through a dynamic tiling step that DeepSeek designed for high-resolution pictures with different aspect ratios. The series comes in three sizes, Tiny, Small and the full DeepSeek-VL2, with 1.0B, 2.8B and 4.5B activated parameters.

This page runs DeepSeek-VL2-Tiny, which DeepSeek built on DeepSeekMoE-3B. DeepSeek lists visual question answering, OCR, document, table and chart understanding, and visual grounding as the jobs the series is meant for.

What you can build

• Alt text and catalog captions. For the sample photo of watermelon slices, the model describes triangular pieces with red flesh and black seeds against a blue background.
• Reading the words in screenshots, signs, product labels or scanned pages.
• Pulling a figure or a header row out of a chart or table inside a report image.
• A review step that flags odd content. Asked what is unusual about a photo of front-loading washing machines, the sample answer points to a person's legs sticking out of a machine door.
• Questions about one subject in a busy frame, for example what a particular person is doing.

Inputs

• One image, uploaded as a file or given as a URL.
• A prompt with your question. It defaults to "Please describe this image in detail.", and the suggested prompts on the page also ask what is unusual in a picture or what a woman in it is doing.
• max_new_tokens, an advanced setting that caps the answer between 10 and 4,096 tokens (default 128).

Outputs

Answers are plain text, saved as a .txt file. The watermelon sample runs to three sentences, while the washing machine answer is a single line. Decoding is greedy, so there is no temperature setting, and the same image with the same prompt should return the same text.

Recommended settings

• For long captions or a full page of OCR, raise max_new_tokens well above 128 so the answer does not stop mid-sentence.
• When you need one detail, ask for it directly, for example "List the column headers in this table".

Limitations

• One image per run. The model itself can take several images in a conversation, but this page accepts a single one.
• Tiny is the smallest of the three sizes, so verify numbers and transcribed text before relying on them.
• Visual grounding is one of the series' tasks, yet you only get text back here. No boxes are drawn on the image.

Safety & compliance

DeepSeek says the DeepSeek-VL2 series supports commercial use under the DeepSeek License Agreement 1.0. That license's use restrictions include harming minors and generating or sharing personally identifiable information without due authorization.