DeepSeek released DeepSeek-V4-Flash-Vision-Exp, a 305B parameter multimodal model under MIT License. Unlike traditional models, it focuses on multimodal agent capabilities, enabling agents to interpret screenshots and execute tasks via tool calls. The release includes full weights and core modules, with community quantized versions available. Performance benchmarks show the model surpassing Claude Opus 4.8 in multimodal agent tasks like ApexBench and Agents' Last Exam, while maintaining strong pure text agent performance. DeepSeek adopted a strategy of launching via API before open-sourcing weights, positioning itself primarily as an API service provider while advancing its complete agent technology stack.