DeepSeek reportedly adds V4 Flash Vision to its API

• DeepSeek’s experimental deepseek-v4-flash-vision-exp is live on its official API, as stevibe’s API announcement reported and DeepSeek’s release note confirms.
• In a form task where the agent saw only an image, it filled every field and submitted after 5 minutes 48 seconds, according to stevibe’s FormSight run.
• The run checked its work before submitting, but check marks landed beside boxes and longer character strings drifted, stevibe’s test report found.
• The FormSight test can be installed as a Bench Pack in BenchLocal, stevibe’s reply said.

DeepSeek’s launch note says the new vision model has out-of-the-box support in DeepSeek Harness 0.1.1. Its vision guide uses OpenAI-compatible content blocks for image input, putting screenshot-reading agents on a familiar API surface.

The API slug is deepseek-v4-flash-vision-exp. DeepSeek labels it experimental and says it retains V4-Flash’s text capabilities for agents, reasoning, and world knowledge in its release note.

The vision documentation lists JPEG, PNG, GIF, and WebP input, and says the service detects a file’s actual content rather than trusting its filename or declared MIME type.

In stevibe’s FormSight run, the model received an image rather than a web page, with no DOM or coordinates. The reported result breaks down cleanly:

• Every field was filled.
• The agent performed a self-check before submit.
• Check marks appeared next to their boxes rather than inside them.
• Longer strings drifted across character boxes.
• The run took 5 minutes 48 seconds, using 288,000 input and 38,000 output tokens.

FormSight is available by downloading BenchLocal and installing the named Bench Pack, stevibe’s reply says. BenchLocal’s documentation describes Bench Packs as installable benchmarks that define their own scenarios, prompts, scoring, verifier lifecycle, and saved run history.