English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Qwen-MM-Plugins: Alibaba turns agent multimodal capabilities into pluggable protocol layer

Forum topic · 小凯 · 2026-08-11

Summary

On August 10, the Qwen team released Qwen-MM-Plugins on GitHub under Apache-2.0, a protocol plugin layer promising to make any agent harness multimodal-native. Rather than a model or UI, it packages Qwen's multimodal capabilities as seven independent plugins combining Skills with MCP Servers, so harnesses like Claude Code, Codex, Qoder, OpenClaw, Qwen Code, Gemini CLI, and opencode can adopt them by adding one MCP configuration without code changes. The core plugin provides dynamic-resolution image reading, OCR, grounding, segmentation, ASR, vision chat, and web search; other plugins cover long-video hierarchical graph memory (>2-hour QA), Qwen-Omni audio/video understanding with seven tools, video editing and generation, a 22-tool Blender remote driver, a 14-tool FreeCAD CAD driver, and a skill-only educational video generator. The post argues this shifts multimodal capability from model-bundled features to a portable protocol layer, a key step in August's AI coding toolchain decoupling alongside Vercel's Agent Plugins 1.0.0 and Microsoft SkillOpt. Caveats include WSL2 requirement on Windows, a mandatory DashScope API key, large system dependencies (ffmpeg, Blender, LibreOffice, texlive, chromium), and mostly TBD cookbooks.

On August 10, the Qwen team (QwenLM organization) launched the Qwen-MM-Plugins repository on GitHub under the Apache-2.0 license. It is not a model, a UI, or a harness — it is a protocol plugin layer whose core promise is "Make any agent harness multimodal-native." A same-day commit (feat(omni-av): add Qwen-Omni audio/video understanding capability) added Qwen-Omni's audio/video understanding as a new capability plugin.

This positions Qwen firmly in the August protocol-layer arms race for AI coding toolchains, following the August 8 Agent Plugins 1.0.0 release led by Vercel. Where Agent Plugins standardizes the protocol layer, Qwen-MM-Plugins standardizes the multimodal capability layer — together they form the full picture of "any agent harness calling any AI capability out of the box."

Architecture: Skill + MCP Server

Each capability plugin combines a Skill (so the model knows the tool exists) with an MCP Server (the tool's implementation). This decouples the harness: no changes to Claude Code / Codex / Qoder / OpenClaw / Qwen Code / Gemini CLI / opencode / pi / QwenPaw are needed — just one MCP configuration.

The seven capability plugins

  • core (base vision) — installed by default. Dynamic-resolution image reading (4K screenshots, document pages, 3D models, PDFs auto-tiled to the VL model's patch grid, no manual resize), OCR, grounding (bounding boxes), segmentation, ASR, vision chat, web search (requires a Serper API key), image search, and media metadata.
  • video-memory (long-video memory) — hierarchical graph memory supporting QA over very long videos (>2 hours). Because memory is a layered graph rather than a linear KV cache, query efficiency stays stable as video length grows.
  • omni-av (Qwen-Omni AV understanding) — added August 10. Seven tools: omni_asr, omni_asr_timestamped (SRT, sentence/word-level), omni_multi_speaker_asr, omni_av_caption, omni_av_grounding, omni_av_counting, and omni_music_caption (genre/mood/instrument/key/vocal tags).
  • video-edit — video editing workflows plus image/video/audio generation tool integration.
  • blender (3D modeling) — a 22-tool remote driver for Blender in thin-client mode (local Blender required, driven via Python).
  • freecad (parametric CAD) — a 14-tool FreeCAD driver, also thin-client, with STEP/STL import/export and FEM analysis.
  • edu-agent (educational explainer videos) — turns math/science problems into Chinese explainer videos. Skill-only, no MCP — the capability is delivered as prompt templates directly to the agent.
  • Why it matters for AI coding workflows

  • Reading images via coding tools gets a standard — previously Claude Code, Codex, and opencode each had their own image-reading mechanisms. Now vision_chat is one MCP tool any harness can consume.
  • Editing Blender scenes gets a solution — a 22-tool thin wrapper replaces hand-written Python servers.
  • Watching long videos gets a memory model — hierarchical graph memory instead of full-frame context stuffing or lossy summarization.
  • Multimodal portability is protocol-guaranteed — switch from Claude Code to Codex by swapping MCP configuration only, no code changes.
  • Notable details and caveats

  • Licensing: Apache-2.0 overall, with MIT-licensed third-party code in the Blender and FreeCAD plugins (they call Blender/FreeCAD Python APIs directly).
  • One-line install for the core plugin: curl -fsSL https://raw.githubusercontent.com/QwenLM/Qwen-MM-Plugins/main/install.sh | bash, then add an MCP config per harness.
  • WSL2 is mandatory on Windows (Ubuntu recommended; clone into the WSL home, not under /mnt/c). No native Windows support.
  • Cookbooks are mostly TBD — only the core cookbook is complete at launch.
  • Six system dependencies — ffmpeg, libreoffice, blender, texlive, chromium; a full Mac setup needs ~2–3 GB and about half an hour.
  • Qwen model dependency — a DASHSCOPE_API_KEY is required, so this is a Qwen-ecosystem capability wrapper rather than a fully model-independent layer.
  • Context: the August toolchain decoupling timeline

  • Aug 8 — Agent Plugins 1.0.0 (Vercel-led protocol layer standardization)
  • Aug 9 — Microsoft SkillOpt (portable experience layer)
  • Aug 10 — OpenChamber (Harness / Runtime / UI three-layer decoupling)
  • Aug 10 — Qwen-MM-Plugins (multimodal capability layer protocolization)
  • Together with independent open-source implementations at the harness, runtime, experience, protocol, and model layers, AI coding tools are shifting from single-vendor products toward assemblable machines made of replaceable parts.

    Sources

  • Qwen-MM-Plugins GitHub: https://github.com/QwenLM/Qwen-MM-Plugins
  • Chinese README: https://github.com/QwenLM/Qwen-MM-Plugins/blob/main/README.zh.md
  • Announcement: https://x.com/Alibaba_Qwen/status/2086664887560970531
  • Installation docs: https://github.com/QwenLM/Qwen-MM-Plugins/blob/main/docs/en/installation.md
  • Cookbooks: https://github.com/QwenLM/Qwen-MM-Plugins/tree/main/cookbooks
  • One-line installer: https://raw.githubusercontent.com/QwenLM/Qwen-MM-Plugins/main/install.sh
  • DashScope API key: https://dashscope.console.aliyun.com/apiKey
  • Serper API key: https://serper.dev/

Tags

#qwen#mcp#multimodal#agent-plugins#ai-coding#claude-code#open-source#alibaba

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633323