All Articles
News

Motion Graphics Without Claude: MiMo V2.6 Made This Short

Michael Leung 2026-10-22 0 views 4 min read
Part 11 of 11 · Local AI on an 8GB GPU

Useful already? ☕ Buy me a coffee and keep these articles free and ad-free.

A desk at night with a monitor showing a street of purple jacaranda trees and glowing abstract bars and circles over the picture

You don't need Claude to make motion graphics videos. I made my Brisbane jacaranda photo guide Short with Xiaomi's MiMo V2.6 Pro instead of Claude Opus 5.5, on a Token Plan that costs about US$6 a month. That plan has no 4-hour usage window and no weekly message cap. The HyperFrames skills hold the recipe, so the model only has to follow it.

On 14 October I wrote about making data videos on an 8GB GPU. Claude Opus 5.5 wrote the animation code in that article. This is the follow-up: what happens when a much cheaper model writes it instead.

What does Opus 5.5 have to do with motion graphics?

It wrote the code. Motion graphics here is not a video model painting every frame. It is HTML, CSS and JavaScript running over pictures, with each frame worked out from the time on the clock. Claude Opus 5.5 turned my figures into counters, bars and slow zooms in the 14 October article, and the browser drew every frame. Numbers stay sharp, and a wrong one is a one-line fix.

That pipeline needed two things: something to write the animation, and something to render the frames into an MP4. Only the first part was expensive.

Can a cheaper model write the same animation code?

Yes. For the Brisbane jacaranda photo guide Short I used MiMo V2.6 Pro, Xiaomi's flagship model, in place of Opus 5.5. It wrote the scene code, I checked it in the browser, and we fixed layout problems the same way I did with Opus: change a line, look again.

Writing HTML animation is ordinary coding work. What matters is whether the model can hold a set of timing rules and follow them. MiMo V2.6 Pro can, and I wrote about its pricing in The Death of Local LLMs? in September. For what the four current Claude models cost instead, see my Claude models comparison.

What are the HyperFrames skills?

They are instruction files you give your coding agent. Rather than the model inventing a video pipeline from scratch, it reads the skills and follows a fixed recipe: build the video as one HTML composition with timing written in data attributes, keep the animation seekable so every frame is repeatable, then let the tool render the MP4.

Browser windows in a row like film frames, each showing abstract shapes and purple flowers, joined by thin light lines

That matters more than the model choice. How to time scenes, move between them and render safely lives in the skill files. It's not the model that knows video; it's the skills. A US$6-a-month model reading a good recipe beats a US$20 one improvising.

The skills also cover motion rules, design and audio, so the agent looks up the recipe instead of guessing. Mine run in OpenCode, and the same skills work in Claude Code and other coding agents.

How was the Brisbane jacaranda Short made?

It is a 9:16 photo guide: my own photos of New Farm Park and UQ Lakes in bloom, with the peak-bloom dates and the rain forecast drawn over them by code. The photos come from my New Farm Park 2026 album and the UQ Jacarandas 2026 page. The forecast figures match the table in my jacaranda photo guide.

A phone standing on a desk, its vertical screen filled with a purple jacaranda canopy and glowing rings and bars over the picture

MiMo V2.6 Pro wrote the animation layer while following the HyperFrames skills. Captions slide in and numbers count up over slowly zooming pictures. Then the renderer turned it into the MP4 on my channel:

What does it cost compared with Claude?

The MiMo Token Plan starts at about US$6 a month. The Claude plans that put Opus 5.5 in your terminal start at US$20 a month, and they come with 4-hour usage windows. Here is the difference at a glance:

  MiMo V2.6 Pro (Token Plan) Claude plans with Opus 5.5
PriceAbout US$6 a month (Lite)US$20 a month
Usage windowsNone, and no weekly caps4-hour usage windows
Models includedMiMo V2.6 Pro and Flash on every tierOpus 5.5 plus the smaller Claude models

The Token Plan is Xiaomi's credit subscription, and four tiers run from about US$6 to US$100 a month. Every tier includes both V2.6 models. Night use is cheaper again: between midnight and 8am Beijing time, credits burn at 0.8 times the normal rate. On the API, Opus 5.5 lists at US$4 in and US$20 out per million tokens, as in my Claude models comparison.

A small silver robot and a large gold robot sitting at opposite ends of a balanced seesaw

Should you use a cheaper model for motion graphics?

For photo guides, data Shorts and anything where text and numbers move over a still picture, yes. The expensive part of this work was never the animation code. Layout problems you can see in the browser before you render, and numbers you can check against your source. Fixing either is a one-line change and a re-render.

Spend the difference on the pictures and the writing. Or don't spend it at all. The whole jacaranda Short cost me a slice of a US$6 plan.

Have you tried a non-Claude model for video work? I'd like to hear which one you used. Send me a note on the contact page.

Sources

Checked on 22 October 2026. This article was drafted with AI assistance and reviewed by me. The article images were made locally with Qwen Image 2.1.

Series: Local AI on an 8GB GPU

Running real AI models on an ordinary 8GB graphics card, and when the cloud is the better deal.

  1. The Death of Local LLMs? How Xiaomi's MiMo V2.6 Just Flipped the AI Economics in Australia
  2. Unlocking Local AI: Running Qwen-Image-2.1 on an 8GB GPU
  3. Local AI Gets a Big Boost: How Ollama's New Jev-Style Decision Models Change Everything
  4. Run Qwen3.8-Flash-Next on an 8 GB RTX 4060 Ti (Here's How)
  5. The Great LLM Debate 2026: Local Hardware vs. Cloud APIs (An Australian Dev's Guide)
  6. EmbeddingGemma 2 Makes On-Device Multimodal Search Work
  7. Underdog Saluki 27B on an 8GB GPU: Why I Got 4.7 Tokens/s
  8. Make Data Videos on an 8GB GPU: Qwen Image + Motion Graphics
  9. MiniCPM-V 4.7 1B on an 8GB GPU: Tested Against 4.6 and 4.5
  10. AI Photo Search on an 8GB GPU: 7,600 Photos Tagged Locally
  11. Motion Graphics Without Claude: MiMo V2.6 Made This Short
Start the series from part 1 →