👀 Seeing is just the beginning. With Qwen-MM-Plugins, turn your favorite agent harness multimodal-native — read images, videos & documents, edit videos, work with 3D/CAD, and more. From multimodal models → multimodal agents. 🚀 Watch it in action:

The entire open source AI community waiting for Qwen-3.8-27b announcement

Hi @Alibaba_Qwen, You may find this useful: It's a proof-of-concept to bring coding agents into Office via ACP. Here is a demo presentation created by Qwen 3.8 Max using OpenCode:

Sadly I do think it works only through api, am I missing something here? Supporter of Qwen and their approach on AI but this? Also Claude Code as a contributor is funny.

The 3D/CAD part is what catches my eye. That’s a much more interesting bridge than just reading another PDF.

This is actually huge 🔥 Finally agents that can properly handle videos, PDFs and even 3D without all the usual frame-by-frame or page-by-page headaches. Qwen-MM-Plugins looking really promising. Gonna try this out today 🚀

i can't handle all these new releases.. we have a multimodal plugins now! so there will be no Qwen 3.8-omni version?

I hope the QwenLM can be used with Qwen 3.8 27B / 35B and can make ai video compilation. I just add all my video from the camera, the ai analyse and understand all the video. Then the ai create a video compilation based on the prompt.

Qwen 3.8 27B when!? We are still waiting 😭

This is where multimodal agents start feeling genuinely useful — not just seeing content, but actually acting on it. Reading videos and documents, editing them, and even working with 3D/CAD opens up a much bigger agent workflow. 🚀

excited for this! great work

This is the exact moment open-source agents stopped playing catch-up. Multimodal-native out of the box. Video edit. 3D/CAD. Long-video memory. Qwen cooked again. 🚀

@Alibaba_Qwen Important correction: my 10.69M-token measurement was during the 50% off-peak window. Normal rates: $6 ≈22.91M/30d $18 ≈91.66M/30d 45.83M / 183.32M are best-case off-peak numbers. Please respond to this.

The 3D/CAD and video-editing plugin hooks are the real signal here. Most enterprises do not need a better chatbot. They need an agent that can manipulate their existing design and media pipelines without rewriting them.

I’m surprised this idea has been untouched for so long

I think people still underestimate how important harnesses are becoming. At some point, the question won’t just be “which model is best?” It’ll be “which model + tools + workflow gives the best result?” Those plugins seems a good step in that direction. What’s the first plugin/tool you’d want in your agent setup?

I like that it’s s giving agents native access to the same visual context humans use, rather than treating images and documents as separate tools.

Multimodal native agents that handle 3D and CAD too, that's a big jump.

love it!

Honestly impressive for a free generation, the motion and lighting nailed it. The one thing giving it away is the resolution, free tier just doesn't have the crispness the paid version probably delivers.🤑

@grok what subs or API service keys do I need to utilize something like this? can this be used to add multimodal capabilities to say - the Reasonix CLI that runs DeepSeek? Or is this focused on QOder type harnesses?

3D/CAD is getting pretty popular recently. Interesting to see what the cost will be for this.

the plugin route is the pragmatic call. most agent harnesses are still text-native with multimodal bolted on through file paths, so making images and video first-class citizens of the loop matters more than another point release of the base model

seeing is just the beginning, but reading the docs is the real adventure

very nicr

This is a big jump in capability!

hey @FreeCADNews you had an appearance here ❤️

Je viens de les installer pour Claude Code. C'est vraiment génial @Alibaba_Qwen.

Open source it

👀📝

@grok can I do this on my Msi 1050 8gen ?

-4k.png 读出每一个数字 @report.pdf 总结第 3 页 @receipt.jpg OCR 并加总金额 @street.jpg 把每辆车框出来 @lecture-2h.mp4 主要观点 + 时间戳（video-memory） @meeting.mp4 带说话人时间戳转写（omni-av） 生成一张小熊猫深夜敲代码的图（video-edit） 建一颗 M6×30 螺栓导出 STEP（freecad） @geometry.png 做成中文讲题视频（edu-agent） --- 9. 和 Hermes 的关系（实用判断） 兼容路径 • 说明: Hermes 有 native MCP：可按 opencode 方式注册 uvx MCP + skill 价值最高 • 说明: core（读图/PDF/视频）+ video-memory + omni-av 注意 • 说明: 依赖 DashScope；部分工具要 ffmpeg；Blender/FreeCAD 要本机 App 重叠 • 说明: Hermes 已有 vision / youtube / video 等 skill；core 的 visualize/动态分辨率读文档更全 若要接到当前 Hermes，最小路径大概是： 1. 装 uv + ffmpeg 2. 配 DASHSCOPE_API_KEY 到 ~/.qwen-mm-plugins/config 3. 在 Hermes config.yaml 加 MCP server：qwen-mm-plugins-core 4. 把对应 skill/ 链到 Hermes skills 目录 --- 10. 学习路径建议 1. 先读 README + docs/en/installation.md + cookbooks/core/usage.md 2. 跑通 core：read_image → visualize(PDF) → ocr/grounding 3. 看 shared/： / / api_dashscope.py / mcp_framework.py 4. 再看一个生成/长视频能力：video-edit 或 video-memory 5. 扩展：照 capabilities/example + docs/how_to_add_new_capability.md 加自己的 cap --- 11. 一句话总结 Qwen-MM-Plugins = 跨 harness 的多模态“外挂操作系统”：用 Skill 教模型用工具，用 MCP+uvx 交付工具，用 DashScope/本地工具补齐读、搜、剪、建、讲全链路。 --- 要继续的话我可以直接帮你做其中一件： 1. 接到 Hermes（写 MCP + skill 配置） 2. 本机装 core 并 verify 3. 深挖某个能力源码（core / video-memory / edu-agent） (2/2)

Multimodal agents become much more useful when perception tools are composable inside the harness, not bolted on as separate apps. The key test will be provenance: can each visual action show which frame, region, or document span drove it?