Qwen-MM-Plugins Is Here: One AI Toolkit for Creators & Developers

There's a specific kind of frustration that comes from working with a coding agent that's otherwise great — until you need it to look at something. A PDF with tables, a screenshot of a broken UI and a receipt you want OCR'd. The agent just can't, so you end up doing that part yourself, then feeding the result back in as text. Half the workflow is still manual. That's what Qwen-MM-Plugins is trying to fix. Not by replacing your agent, but by giving it the ability to actually process visual inputs on its own. Also Qwen-MM-Plugins showed up a few days ago with the Qwen3.8-Max release.

I ran into this personally while building a PDF analyzer. Documents with complex column layouts, embedded tables, mixed content — my agent wrote clean code, but had no idea what was in the files it was supposed to process. After plugging in the core capability from Qwen-MM-Plugins, those same documents came through properly structured. No preprocessing on my end, no manual extraction step. That chunk of work just stopped existing.

What This Thing Actually Is

Qwen-MM-Plugins isn't a model or a standalone tool. It's a set of capabilities you bolt onto whichever coding agent you're already using. Qwen Code, Claude Code, Codex, OpenClaw, and Gemini CLI all supported. Install a capability, and the agent starts using the right visual tool automatically when you reference a file. Six capabilities ship in the repo, each installed on its own:

core

The one you'll almost certainly want. Point it at an image, a video clip, a PDF, or a 3D model file and it figures out the right level of detail to process it at — a tiny thumbnail and a massive 4K file both get handled properly without you adjusting anything first. Beyond basic reading, core also does OCR, lets you locate specific objects within an image, audio transcription, and web search.

Install name: qwen-mm-plugins-core

video-memory

Useful when you're dealing with long recordings. A two-hour lecture, a full product demo, a recorded meeting — it builds an internal memory structure over the entire video so you can ask specific questions and get answers tied to actual timestamps. Without something like this, long video content is basically unusable in an agent workflow.

Install name: qwen-mm-plugins-video-memory

video-edit

Covers video editing workflows alongside image, video, and audio generation. If your work involves producing or modifying media files, this is the one.

Install name: qwen-mm-plugins-video-edit

blender

This one's interesting. It connects to a running Blender session and drives it through Python — 22 tools for modeling, materials, lighting, and rendering. The agent doesn't replace Blender, it just controls it. So you can describe what you want modeled and the agent handles the Blender side.

Install name: qwen-mm-plugins-blender

freecad

Same concept, different software. FreeCAD with 14 tools, parametric modelling, STEP and STL file handling, FEM analysis included. Good fit if you're doing any mechanical or engineering design work and want your agent in that loop.

Install name: qwen-mm-plugins-freecad

edu-agent

Takes a math or science problem typed or as an image and produces a step-by-step Chinese explainer video or an interactive page from it. Worth noting this one has no MCP server; it's a skill-only setup.

Install name: qwen-mm-plugins-edu-agent

How It Works Under the Hood

Each capability is two pieces. A skill registers the toolset with the model so it knows these options exist. An MCP server actually runs those tools when called. The MCP server spins up through uvx the first time it's needed — no separate pip install, it manages its own dependencies.

Getting It Installed

The quickest path is the guided installer:

curl -fsSL https://raw.githubusercontent.com/QwenLM/Qwen-MM-Plugins/main/install.sh | bash

Run this once. It covers everything — initial setup, configuration, checking what's working, and removal if needed. It also writes a config file to ~/.qwen-mm-plugins/config that every harness (both terminal and GUI-based) reads from automatically. Set it up once, done.

Windows: WSL2 is the only option right now. When you clone the repo, do it inside your WSL home directory like ~/code — not under /mnt/c or any mounted Windows path. Native Windows support isn't there yet.

If you'd rather install per-harness manually, here's each one (replace core with whichever capability you're installing):

Qwen Code

qwen extensions install https://github.com/QwenLM/Qwen-MM-Plugins.git:qwen-mm-plugins-core --consent

Claude Code

claude plugin marketplace add https://github.com/QwenLM/Qwen-MM-Plugins.git
claude plugin install qwen-mm-plugins-core@qwen-mm-plugins

Codex

codex plugin marketplace add https://github.com/QwenLM/Qwen-MM-Plugins.git
codex plugin add qwen-mm-plugins-core@qwen-mm-plugins

OpenClaw

openclaw plugins install qwen-mm-plugins-core --marketplace https://github.com/QwenLM/Qwen-MM-Plugins.git

System Dependencies

FFmpeg is required for anything involving video or audio. The others — LibreOffice, Blender, TeXLive, Chromium — are only needed depending on which plugins you're using. Check what's present on your machine:

bash install.sh verify

This also confirms your API keys are set and flags anything missing.

API Keys

Two keys cover everything:

• DASHSCOPE_API_KEY — required for vision chat, OCR, grounding, transcription, generation, and video-memory features
• SERPER_API_KEY — required for web search and image search

Drop them into your shell environment or write them to ~/.qwen-mm-plugins/config. That file gets picked up by GUI harnesses as well as terminal ones. Or just run:

bash install.sh configure

Day-to-Day Usage

After setup, you reference a file with @ and describe what you need. The agent decides which tool handles it:

@dashboard.png - Pull out every number visible in this chart.
@report.pdf - What does page 3 say about Q2 revenue?
@receipt.jpg - OCR this and add up the line items.
@street.jpg - Find and box every vehicle in this image.
@meeting-recording.mp4 - Summarize the key decisions with timestamps.

No need to specify which tool the model reads the context from and picks.

Summary

If you're someone who hits the image/document wall regularly with your coding agent, the core plugin alone is probably worth the setup time. It covers most of what comes up in day-to-day PDFs, screenshots, OCR, and visual grounding. The Blender and FreeCAD capabilities are genuinely useful if you're in that niche. Agent-controlled 3D modelling isn't something most tools offer at all, let alone in a form that connects to software you're already running.

GitHub: github.com/QwenLM/Qwen-MM-Plugins