Qwen has published Qwen-MM-Plugins, an open source project that makes multimodal capabilities available as a modular tool layer for AI agents.
Instead of connecting an agent to one model expected to handle every visual and audio task, the project lets the agent discover and invoke separate capabilities when they are needed.
What the tools can do
The core capability supports image, video, document and 3D-model reading, as well as OCR, object grounding, segmentation, audio transcription and vision chat. Other capabilities add long-video memory, video editing and generation, and integrations for Blender and FreeCAD.
Each capability is installed separately. It includes a skill that describes the available tools and, when needed, an MCP server that provides the tool interface.
A longer multimodal workflow
An agent can read an image, locate an object, crop the relevant area, run OCR, compare the result with a video frame, transcribe the audio and combine the evidence into one answer.
This changes the architecture: multimodality becomes a set of capabilities that can be replaced, extended and chained inside an agent harness. The repository documents support for Claude Code, Codex, Qwen Code and Gemini CLI.
The hard part moves up the stack
Modularity does not automatically solve reliability. The agent still needs to choose the right tool, validate its output, manage permissions and know when to stop instead of continuing autonomously.
The practical question is changing. It is less about which model can do everything, and more about which tools an agent can select, invoke and chain safely.
See the official Qwen-MM-Plugins repository on GitHub for the capability list and installation details.