DeepSeek Harness Updates Multimodal Support, Hinting at Future V4 Vision Model
DeepSeek Harness has updated to support multimodal inputs, allowing its model adapter to configure native image requests and enabling commands like `/goal` and `/plan` to accept images. The update also allows users to reference screenshots, local files, and historical conversations simultaneously in a single query. This update prepares the ecosystem for a potential future DeepSeek V4 vision model while allowing developers to use third-party vision models in the interim. It significantly enhances the capability of AI agents to handle complex, image-based workflows and multi-turn multimodal interactions. The update introduces persistent image attachments for MCP and ACP, fixes bugs related to large image sizes, and separates Claude Code and Codex into on-demand Profile Bundles. Additionally, the SQLite backend received performance improvements but introduced an incompatible data structure.
## BACKGROUND
DeepSeek Harness is an open-source agent runtime developed by DeepSeek AI, designed with a modular architecture where capabilities like models and tools are treated as swappable plugins. It provides the infrastructure for developers to build, test, and deploy AI agents that can interact with various environments.