300 motion templates packed into DeepSeek, video starts becoming a coding problem
The OpenMotion open-source plugin is integrated into DeepSeek Harness, breaking reference videos into shots, rhythm, and transitions, then letting a local Agent reconstruct them as HTML scenes. The number of templates is the entry point; Motion.md is the method it wants to promote.
Tom Huang announced on X that the OpenMotion open-source plugin is live on DeepSeek Harness. Strictly speaking, what got stronger is DeepSeek Harness's plugin ecosystem, not that the DeepSeek model suddenly knows how to render video. OpenMotion is an open-source project developed by OpenDesign under the Apache-2.0 license, and the repository was renamed from nexu-io/html-anything.
Tuola Ge used OpenMotion to make a recreation video, creating a segment each for Gemini Tasks, dots, Claude Duet, Grok Bot, and Muse. In the tweet, he used Gemini 4 Neon, Opus 5.5, and Fable 5.2 as references, letting readers judge how close it is.
OpenMotion takes a more engineering-oriented approach: write video as web pages. Scenes, layout, transitions, and timing are described with HTML, CSS, JavaScript, and GSAP; screen recordings, generated clips, images, music, and voiceovers are all put into the same project as assets. The same project can show a timeline in Studio and render to MP4.

What it aims to solve is recreation and revision. Upload a reference video or paste a social link; the Agent first breaks down shot order, transitions, and audio cues, marks replaceable assets, then rebuilds HTML scenes using your brand, copy, and assets. Motion.md records shot order, timing, transitions, and replaceable assets. The materials mention 300+ motion DESIGN.md files, 152+ design systems, and 100+ product promo templates from top companies. Tom Huang's figures are 470,000 lines of code in 14 days, supporting 20+ agents including Codex, Claude Code, and DeepSeek Harness. The README lists 22 local agents.

There are two installation paths. It can be installed with one click from the ChatGPT Marketplace. You can also run npx skills add nexu-io/openmotion --skill openmotion in a local Agent, then have the Agent read the installation docs and connect to local Studio and MCP. OpenMotion supports BYOA and BYOK: use your existing Agent, and use your own model API key. The local workflow doesn't charge a software subscription fee or per-render fees; models and third-party services are billed through your own accounts. Source preview and export depend on Node.js 26.11.1, Chrome/Chromium, FFmpeg/FFprobe.

A user asked whether only DeepSeek models can be used. Tuola Ge replied that any can be used, and OpenMotion can be invoked in 20+ agents. Someone reminded that if brand templates are used commercially, it's best to check how licensing works first. The materials don't give an answer; confirm before use. Another commenter said more templates is good, but it depends whether the output videos feel template-y. That concern isn't unnecessary. 300 templates can solve the starting point, but also easily make every video look the same. OpenMotion puts the modification entry points in source code and Motion.md; whether it can be modified to feel on-brand still depends on the person.
If you want to try it, start with the minimal workflow: choose a template, click Recreate template, replace the brand and copy, align the timing in Studio, and export MP4. Don't rush into batch production right after installing. First look at how one video goes from reference to finished cut: which parts rely on the template and which rely on your judgment.
Open-source repo: https://github.com/nexu-io/openmotion
Tom Huang's original tweet: https://x.com/tuturetom/status/2123367688655483114
Tuola Ge's original tweet: https://x.com/Tuola0205/status/2123362270139109827
发布时间: 2026-11-19 21:30