DeepSeek Open-Sources V4-Flash-Vision-Exp for AI Agents

DeepSeek has released the weights for its first multimodal model, V4-Flash-Vision-Exp, as open source. The 305-billion-parameter model is available on Hugging Face under the MIT License.

The open-weights release went live on August 31, following an API-only debut on DeepSeek’s platform on August 21. The package makes the model available beyond DeepSeek’s hosted API and completes the staged rollout.

V4-Flash-Vision-Exp Is Built for Agentic Vision Tasks

V4-Flash-Vision-Exp is designed for AI agents that need to work with visual information as part of a task. Rather than focusing on conventional visual question answering, it is intended to read web screenshots, software interfaces, and charts, then support task execution through tool calls.

DeepSeek describes the model’s multimodal agent capabilities as being “close to Opus-4.8,” Anthropic’s flagship model. Its stated purpose is not simply to interpret images for a chatbot conversation, but to provide a visual layer for agents that interact with digital environments.

That distinction matters for workflows where an agent needs to recognize what appears on a screen before taking its next action. The model is positioned to process visual inputs that arise while navigating websites, interacting with software, or interpreting chart-based information.

DeepSeek V4-Flash-Vision-Exp Benchmark Results

The model’s Hugging Face page includes benchmark results comparing V4-Flash-Vision-Exp with Opus-4.8 across agent-focused evaluations.

On ApexBench Pass@1, V4-Flash-Vision-Exp scored 36.5, compared with 39.4 for Opus-4.8. On Agents’ Last Exam, DeepSeek’s model reached 27.3, ahead of Opus-4.8’s 25.7.

Text-only agent performance also remained steady after visual capabilities were added. On Terminal Bench 2.1, the score increased from 82.7 to 83.9.

The results were not uniformly favorable. On NL2Repo, V4-Flash-Vision-Exp scored 57.7, while Opus-4.8 scored 69.7. The published figures show that the model performs differently across benchmarks rather than matching or exceeding Opus-4.8 in every category.

DeepSeek Vision Model Pricing and Response Speed

An early comparison from The New Stack tested V4-Flash-Vision-Exp against Google’s Gemini 3.7 Flash on nine vision tasks. The comparison found that the two models matched on accuracy in those tasks.

The cost figures differed sharply. DeepSeek’s model had a reported total cost of $0.0039, compared with $0.0122 for Gemini 3.7 Flash. That put the DeepSeek result at roughly one-third of the cost in that comparison.

However, the DeepSeek model took more than twice as long to respond. So the comparison points to a clear tradeoff: lower reported cost, with slower response time.

OpenRouter lists V4-Flash-Vision-Exp pricing at:

• $0.22 per million input tokens
• $0.66 per million output tokens

For comparison, Gemini 3.7 Flash is listed at:

• $0.75 per million input tokens
• $3.75 per million output tokens

What the Open-Source Release Includes

The open-source package contains the components needed to work with DeepSeek V4-Flash-Vision-Exp beyond its API offering:

• Model weights
• Tokenizer
• Prompt encoding references
• A minimal PyTorch inference implementation

Community-built quantized versions for llama.cpp, LM Studio, and Ollama appeared within hours of the release. These versions expand the set of available deployment options around the model.

The release uses an MIT License, and the model weights are hosted on Hugging Face under the DeepSeek AI account.

How V4-Flash-Vision-Exp Fits DeepSeek’s Agent Toolkit

DeepSeek introduced V4-Flash-Vision-Exp as an API model first, then made its weights available later. This sequence matches the company’s broader “API first, open weights second” approach.

Alongside the API debut on August 21, DeepSeek launched Harness 0.1.1 for continuous task execution. The vision model can now serve as the visual or perceptual component within that wider agent toolkit.

Together, the releases connect task execution with a model designed to interpret screenshots, interfaces, and charts. The approach gives DeepSeek’s agent tooling a multimodal layer that can process visual information before an agent makes tool calls.