Qwen plugin pack gives your coding agent eyes, hands, and video memory

Qwen has shipped an open-source plugin pack that extends coding agents with vision, audio, and 3D skills. Qwen-MM-Plugins, published on GitHub under the QwenLM organization, installs into Claude Code, Codex, Qoder, OpenClaw, Qwen Code, and Gemini CLI. The license is Apache-2.0, and the repo states its goal in one line: make any agent multimodal-native.

Each capability installs as a skill, so the model knows the toolset exists, plus an optional MCP server that provides the tools themselves, launched on demand by uvx. The first capability, called core, handles foundational vision: dynamic-resolution reading of images, videos, documents, and 3D models, plus OCR, grounding, segmentation, ASR, vision chat, and web search. The other packs are video-memory, omni-av, video-edit, blender, freecad, and edu-agent.

What ships in the new Qwen repo

The seven packs install separately, and the repo documents what each one brings.

The repo ships cookbooks showing each capability in action, with tool listings, setup steps, and worked cases. The tool counts show how granular the integration gets: 22 tools for modeling, materials, lighting, and rendering in Blender; 14 for parametric modeling and FEM analysis in FreeCAD.

Install once, use across Claude Code, Codex, and the rest

Qwen recommends a guided installer, a single script that handles install, configure, verify, and uninstall across every supported harness. It drives each harness's own native install under the hood and writes one shared config file at ~/.qwen-mm-plugins/config, so terminal and GUI harnesses read the same setup. The one-liner pulls from install.sh in the repo.

Manual install works too. Harnesses with plugin marketplaces, which the repo lists as Claude Code, Qoder, Codex, OpenClaw, and Qwen Code, take a capability by adding the GitHub repo as a marketplace and installing a named pack. Qoder is Alibaba's own agent IDE; its 1.0 release claimed a 40% cut in agent input tokens, a test Alibaba ran itself (see the details). Other harnesses, including Gemini CLI, opencode, pi, and QwenPaw, register the skill and MCP server in their own config. The repo notes that the easiest path is to just ask the agent to install the capability. Windows users get one supported route, WSL2; native Windows has not yet been validated.

What it takes to run

Dependencies are deliberately light. uvx installs the Python dependencies for the chosen profile on first launch, so there is no manual pip. The only things you install yourself are system tools: ffmpeg for video and audio, plus optional libreoffice, blender, texlive, or chromium for visualization. A bash install.sh verify command self-tests the setup, confirms the API key, and reports missing system tools.

Native reading does not need an API key. The API-based tools do: DASHSCOPE_API_KEY covers vision chat, OCR, grounding, transcription, omni audio-video understanding, generation, and video-memory build; SERPER_API_KEY covers web search, web extractor, and image search. The guided installer's configure step writes them to the shared config file, which is read whenever a variable is not already in the environment. DashScope is also where Alibaba's Wan3.0-Video charges by the second, $6 for a 30-second 1080P clip (the cost breakdown), something to budget for if you plan to use the video-edit pack.

In practice, reading is dynamic-resolution: every image, video frame, and document page is auto-scaled to the VL model's patch grid. The repo's examples include reading every number in a 4K dashboard screenshot, OCR-ing a receipt and totaling line items, drawing boxes around every car in a street scene, asking for the main points of a two-hour lecture with timestamps, counting completed passes in a sports clip, and modeling an M6 hex bolt and exporting it as STEP. Those two cases are exactly the watching, not making, territory that Perceptron Mk1 is betting on.

Alibaba Cloud is betting that multimodal workflows belong outside the Qwen ecosystem alone. The direction matches Qwen Studio's work on chatbot, image and video understanding, document processing, and web search integration, and Qwen2.5-Omni, the end-to-end multimodal model family that handles text, images, audio, and video in a stream. Qwen2.5-Omni outperformed Gemini-1.5-Pro on OmniBench and fits under 12GB (full results). With Qwen-MM-Plugins, the same capabilities now ride on top of other vendors' agent harnesses. The license is Apache-2.0, with third-party MIT-licensed code vendored for the Blender and FreeCAD capabilities, and NOTICE files included for attribution. The open-weights streak extends to Alibaba's biggest model, Qwen3.8-Max, which coded alone for 16 days (our coverage).