DeepSeek Opens a Vision Experiment Model for Its V4 Series

Introduction

DeepSeek is extending its V4 family beyond text reasoning and code generation. The company has made deepseek-v4-flash-vision-exp available through its API platform, providing image understanding for the first time through an official V4 interface. Developers can now place images alongside text in requests and use the model inside agents, automation workflows, and coding tools.

Key points

• Official vision input is now available: The model accepts mixed image-and-text prompts.
• Several image delivery methods are supported: Developers can use inline base64 data, external URLs, or the Files API.
• Multiple API formats are covered: Chat Completions, Messages, and Responses are all supported.
• Images are billed as tokens: Each image can consume up to 384 tokens, with pricing aligned with V4-Flash.
• Agent benchmarks mostly improve: Five of seven listed evaluations rise against V4-Flash-0731, while one declines slightly.

The benchmark pattern suggests that Vision-Exp may involve more than attaching a visual encoder to an existing language model. Toolathlon-Verified increases from 70.3 to 75.9, while DeepSWE moves from 54.4 to 59.3. Both tests are closely related to tool use, codebase modification, or extended execution. That combination points to possible improvements in how the model interprets an operating environment, plans a task, and coordinates external tools.

Other changes are also positive but more moderate. Terminal Bench 2.1 rises from 82.7 to 83.9, NL2Repo from 54.2 to 57.7, DSBench-Hard from 59.6 to 63.6, and AutomationBench from 25.1 to 25.7. Cybergym falls from 76.7 to 75.3, showing that the update is not uniformly better across every task.

Why it matters

For developers, the immediate benefit is a simpler path to multimodal agents. Screenshot analysis, interface understanding, document inspection, and vision-assisted coding can be handled through the same general API surface instead of requiring a separate vision service. The image token limit and pricing alignment also make initial cost planning easier.

The comparison with Opus 4.8 nevertheless calls for a measured interpretation. Vision-Exp reaches 59.3 on DeepSWE, above Opus 4.8’s 58.0, and scores 75.9 on Toolathlon, only 0.3 points below Opus. However, it trails by 12 points on NL2Repo and 8.1 points on DSBench-Hard. The most defensible conclusion is that the model is approaching leading systems and is competitive on selected coding-agent tasks, not that it has surpassed them generally.

The “Exp” suffix remains important. DeepSeek still labels the model experimental, so production users will need to evaluate visual accuracy, reliability on complex images, API stability, and real-world cost. Even with those qualifications, the release fills a notable gap in the V4 product line. It places visual understanding and agent execution on the same API path, potentially making DeepSeek more practical for multimodal automation.

InfoQ 中文