XNews

Qwen-MM-Plugins Multimodal Agent Toolkit

Qwen-MM-Plugins Multimodal Agent Toolkit

Qwen team released Qwen-MM-Plugins, a set of native multimodal plugins that integrate image, video, document, audio, and 3D modeling capabilities into any agent harness via MCP. - The architecture separates each ability into a Skill (a SKILL.md prompt telling the model when to use the tool) and an optional MCP server pulled on first call. - Seven modules: Core (vision, OCR, grounding, segmentation, ASR, search), Video-Memory (hierarchical memory for 2+ hour videos), Omni-AV (audio-video with timestamps and speaker labels), Video-Edit (generation and editing), Blender (3D modeling via Blender thin client), FreeCAD (parametric CAD with FEM), and Edu-Agent (Math/science problem explanation video generation). - This allows pure text agents to become multimodal-native without retraining.

Key insights

Modular MCP-Based Approach Simplifies Multimodal Agent Upgrades

By decoupling tools from agent code and using on-demand MCP servers, Qwen-MM-Plugins lets any text agent instantly acquire vision, audio, and 3D skills without modifying its core architecture. The Skill definition in markdown also makes it transparent when the tool should be invoked.

Source posts