Chinese AI company DeepSeek has released V4-Flash-Vision-Exp, an experimental multimodal model that adds image understanding to its text capabilities. On the company's own benchmarks, it nearly matches Opus 4.8 on agent tasks.

According to DeepSeek, the new version extends V4-Flash with image processing while keeping the base model's text performance in reasoning and world knowledge. That distinction matters: adding multimodal capability usually costs something on the text side.

The target is agent workflows, not chat

The company positions the model explicitly for agent-based applications. It is designed to work with different agent frameworks and to combine visual understanding with tool use. In practice it can:

  • Describe images — explain what a picture contains.
  • Extract text from screenshots — read the writing in interface captures.
  • Analyze diagrams — interpret charts and schematics.

The model handles JPEG, PNG, GIF and WebP. It determines the format from the actual file content rather than the filename or the declared MIME type. That sounds like a small detail but it matters in agent workflows, where the extension on a file an agent has downloaded is often not trustworthy.

The model works with OpenAI's Chat Completions and Responses APIs and with Anthropic's Messages endpoint. DeepSeek also released a version of its own Harness framework that supports the new model out of the box.

The pricing is the real story

There are three ways to send an image: embed it directly with Base64 encoding, point to a publicly accessible URL, or use the new, free Files API. The Files API lets a file be uploaded once and referenced by ID across multiple requests.

The striking detail sits on the cost side. An optional field downscales images when fine visual detail is unnecessary, saving tokens. The model normalizes images to roughly 800 by 800 pixels depending on aspect ratio before processing. As a result, whatever the original resolution, each image costs at most 384 tokens. Pricing follows V4-Flash rates, so vision carries no separate charge.

A fixed ceiling makes the cost of image-heavy agent workflows predictable in advance. Variable token consumption per image had been one of the main reasons an agent working in a screenshot loop produced an unpredictable bill.

The limits

A single request can include up to 600 images. Maximum edge length is 8,192 pixels per side, but that drops to 4,096 once a request contains 15 or more images. Images can only go in user messages. The "experimental" in the model's name is worth keeping in mind too: the comparison figures come from the company's own internal benchmarks, with no independent verification yet.