DeepSeek has launched DeepSeek-V4-Flash-Vision-Exp on Hugging Face, the first experimental multimodal model in the V4 family, released under the MIT License. The model has 305B parameters and is based on the V4-Flash-0731 foundation. It integrates a visual encoder and an Aligner into the original language model, gaining image understanding capabilities through continuous training.

image.png

Not "Describing Pictures," Focused on Multimodal Agent

Differing from traditional multimodal models that focus on visual question answering, DeepSeek has set a new direction for this Vision version: the official emphasizes the multimodal agent capability, allowing agents to directly read visual information such as web screenshots, software interfaces, and charts, and then execute tasks by calling tools. In other words, this "eye" is not for human use but for the agent.

The open-source content is quite complete, including model weights, Tokenizer, Prompt Encoding reference implementation, and a minimal PyTorch inference implementation, covering core modules such as the visual encoder, Aligner, DFlash Attention, MoE, and Hyper-Connections. The community responded quickly, with seven quantized versions already available on the Hugging Face page for llama.cpp, LM Studio, and Ollama.

There are also notable details in the timeline: On August 21, the model was first launched on the DeepSeek API, at that time only accessible via API without providing weights. Developers could input text and images simultaneously through the API, with images charged by token. Ten days later, the weights were officially opened, making local deployment and secondary development possible. This "sell API first, then open source" approach also reveals DeepSeek's positioning as primarily an API service.

image.png

Performance Approaching Claude Opus 4.8, Official Language Remains Moderate

From the benchmark results, the pure text agent capabilities after adding visual abilities remain largely unaffected: Terminal Bench 2.1 increased from 82.7 to 83.9, and DeepSWE increased from 54.4 to 59.3, even surpassing Opus 4.8's 58.0. The improvement of the multimodal agent is more significant: ApexBench Pass@1 reached 36.5, Agents' Last Exam achieved 27.3, exceeding Opus 4.8's 25.7, and ZeroBench Pass@5 reached 35.0, surpassing the opponent's 34.0.

However, DeepSeek did not claim "comprehensive superiority": on the NL2Repo project, its score was 57.7, far behind Opus 4.8's 69.7. The official statement remains moderate, stating only that "multimodal agent capabilities are close to Claude Opus 4.8." Looking at recent updates, DeepSeek is assembling a complete agent technology stack: the model handles reasoning and tool calls, Harness manages continuous task execution, and Vision allows agents to directly read visual information from the computer. This open-source release is a key move in completing this puzzle.