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

Qwen-MM-Plugins: Alibaba Open-Sources a Pluggable Multimodal Protocol Layer for Any Agent Harness

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-level layer that makes any agent harness multimodal-native without modifying harness code. The repository combines Skills and MCP servers to expose Qwen's multimodal capabilities as pluggable modules. Seven capability plugins are included: core (vision, OCR, grounding, segmentation, ASR, web/image search), video-memory (hierarchical graph memory for >2 hour videos), omni-av (Qwen-Omni audio/video understanding with 7 tools including timestamped ASR, multi-speaker separation, AV grounding, counting, and music captioning), video-edit, blender (22 tools, thin-client mode), freecad (14 tools for parametric CAD), and edu-agent (a pure Skill with no MCP). Installation is a one-line script that wires the plugins into Claude Code, Codex, Qoder, OpenClaw, Qwen Code, Gemini CLI, and opencode. Key constraints: WSL2 is required on Windows, ~2-3 GB of system dependencies (ffmpeg, libreoffice, blender, texlive, chromium) must be installed, and a DashScope API key is mandatory, making the stack Qwen-model-tied. Qwen-MM-Plugins complements the August 8 Agent Plugins 1.0.0 protocol standardization by adding a standardized multimodal capability layer, completing the "any harness times any capability" stack.

Overview

On August 10, the Qwen team (QwenLM) released Qwen-MM-Plugins on GitHub under Apache-2.0. It is not a model, not a UI, and not a harness; it is a protocol plugin layer that makes any agent harness multimodal-native. The core promise is *Make any agent harness multimodal-native*. A same-day commit added Qwen-Omni audio/video understanding as a new capability plugin (feat(omni-av): add Qwen-Omni audio/video understanding capability).

The release fits into the August protocol-stack arms race, complementing the August 8 Agent Plugins 1.0.0 (led by Vercel). Where Agent Plugins 1.0.0 standardizes the protocol layer, Qwen-MM-Plugins standardizes the multimodal capability layer; together they complete the picture of "any harness can call any AI capability out of the box."

Architecture: Skill + MCP Server

Each capability plugin = a Skill (so the model knows the tool exists) + an MCP Server (the tool implementation). The benefit is harness decoupling: Claude Code, Codex, Qoder, OpenClaw, Qwen Code, Gemini CLI, opencode, pi, and QwenPaw need no code changes — just an MCP configuration.

Seven Capability Plugins

  • core (default-installed, base vision) — dynamic-resolution image reading (4K screenshots, documents, 3D models, PDFs auto-split into VL-model patches), OCR, grounding (bounding boxes), segmentation, ASR, vision chat, web search (Serper API key required), image search, media-info metadata.
  • video-memory (long-video memory) — hierarchical graph memory supporting >2 hour video QA. Memory is a hierarchical graph, not linear KV cache, so query efficiency stays stable as video length grows.
  • omni-av (Qwen-Omni audio/video, new August 10) — 7 tools: omni_asr (plain text transcription), omni_asr_timestamped (SRT sentence/word timestamps), omni_multi_speaker_asr (multi-speaker separation), omni_av_caption (time-segment descriptions), omni_av_grounding (text query to event timestamp), omni_av_counting (event/action/object counts), omni_music_caption (genre/mood/instrument/key/vocal tags).
  • video-edit — editing workflows plus image/video/audio generation integrations.
  • blender (3D modeling) — 22 tools, thin-client mode (local Blender required, driven via remote Python).
  • freecad (parametric CAD) — 14 tools, thin-client mode, STEP/STL import-export and FEM analysis.
  • edu-agent (instructional videos) — converts math/science problems into Chinese explainer videos. Pure Skill mode, no MCP — capability is delivered as agent prompt templates.
  • Why This Matters for AI Coding Workflows

    The trajectory of multimodal capability delivery:

  • Past 12 months — multimodal features were model-native (Qwen-VL, GPT-4V, Gemini 1.5 Pro, Claude 3.5 Sonnet). Switching models meant rewriting tool-calling code.
  • Past 6 months — Anthropic's MCP protocol (November) began standardizing tool calls, but multimodal capability remained model-bound.
  • Qwen-MM-Plugins — splits Qwen's multimodal capabilities into 7 independent plugins, each exposed via MCP. Harnesses add one MCP config to call Qwen's multimodal tools without being locked to the Qwen model.
  • Concrete wins for AI coding:

  • "AI coding tool reads images" is now standardizedvision_chat works in any harness via MCP, replacing Claude Code's Read tool, Codex's vision API, and opencode's file-include mechanism.
  • "AI coding tool edits Blender" is now solvable — 22 thin Blender wrappers replace custom Python servers.
  • "AI coding tool watches long videos" now has a memory model — hierarchical graph memory replaces "dump all frames" or "summarize" heuristics, with log-scaling rather than linear growth.
  • Multimodal capability portability is now protocol-guaranteed — switch from Claude Code to Codex without code changes; only the MCP config changes.
  • Practical Notes and Constraints

  • Apache-2.0 main, MIT exception — Blender and FreeCAD plugins include MIT-licensed third-party code (for direct Blender/FreeCAD Python API calls); all others are Apache-2.0.
  • One-line installcurl -fsSL https://raw.githubusercontent.com/QwenLM/Qwen-MM-Plugins/main/install.sh | bash for the core plugin; each harness adds its own MCP config.
  • WSL2 is mandatory on Windows — the docs explicitly require WSL2 (Ubuntu recommended); clone inside the WSL home directory, not on /mnt/c mount points. No cross-platform out-of-the-box support.
  • Cookbooks largely TBD — only the core cookbook is complete; others are marked "TBD," so the August 10 release mainly demos core capabilities.
  • System dependenciesffmpeg, libreoffice, blender, texlive, chromium. A full install requires ~2–3 GB and ~30 minutes on macOS.
  • DashScope API key is required — Qwen-MM-Plugins is not an independent capability layer; it is a capability wrapper around the Qwen model ecosystem. Serper API key is optional (for web search).
  • August AI Coding Tool-Chain Timeline

  • 8-08 Agent Plugins 1.0.0 — Vercel-led protocol-layer standardization.
  • 8-09 Microsoft SkillOpt — experience-layer portability.
  • 8-10 OpenChamber — harness / runtime / UI three-layer decoupling.
  • 8-10 Qwen-MM-Plugins — multimodal capability-layer protocolization.
  • August 10 marks the "full-stack decoupling" moment for AI coding tooling: harness (opencode), runtime (OpenChamber), capability layer (Qwen-MM-Plugins), experience layer (SkillOpt), protocol layer (Agent Plugins 1.0.0), model layer. Each layer now has an independent open-source replacement, turning "AI coding tools" from single-vendor products into assemblies of swappable parts.

    Sources

  • Repository: https://github.com/QwenLM/Qwen-MM-Plugins
  • Chinese README: https://github.com/QwenLM/Qwen-MM-Plugins/blob/main/README.zh.md
  • Alibaba_Qwen announcement: https://x.com/Alibaba_Qwen/status/2086664887560970531
  • Installation docs: https://github.com/QwenLM/Qwen-MM-Plugins/blob/main/docs/en/installation.md
  • Cookbooks index: 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-mm-plugins#mcp-protocol#agent-harness#multimodal-ai#plugin-architecture#model-context-protocol#ai-coding-tools#open-source

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