Math Videos Rendered 90x Faster

ManimGX is a 3D animation engine meant for AI agents to drive. It keeps the Python API from Manim Community Edition but swaps the renderer for one written in Rust on wgpu. The problem it targets is easy to recognize if you've had an agent generate Manim videos. Rendering is slow, LaTeX installs are a pain, and the agent can't tell whether its scene looks right. ManimGX goes after all three.

How it works

Your agent writes ordinary Manim CE style Python. The README says it's statically typed, with keyword arguments and .animate chains. The Rust and wgpu engine then renders it to MP4. 2D and 3D work in the same scene, so surfaces, meshes, moving cameras, text, equations and plots can share a frame.

Text and math go through Typst, and LaTeX input gets translated with mitex. That means you don't need a LaTeX install at all. Video is encoded with x264. On Linux machines without a usable GPU, the wheels bundle Mesa's lavapipe so Vulkan rendering still runs on the CPU.

For agents, the most useful piece is manimgx inspect. It produces storyboards and 2D layout checks, which gives the model something concrete to look at before it renders the whole video. There's also self.say() for syncing speech, plus a browser build that runs on Pyodide and WebGPU (npm install manimgx).

The speed claim: 36.2 seconds of 1080p60 video across five scenes, rendered in 3.08 seconds on an M4 Pro. The README puts that at 90.6x faster than ManimCE, 11.5x faster than ManimGL and 18.7x faster than Blender Workbench.

Try it

pip install manimgx
manimgx render scene.py

You need Python 3.13+ on macOS, Linux or Windows. Or you can paste Follow https://manimgx.academa.ai/llms.txt. into Claude Code, Codex or Cursor and let the agent do the setup.

My take

I like that it's built around the agent loop. Fast renders, no LaTeX dependency and an inspect command that returns layout checks are the things that actually slow down agent-made explainer videos. Keeping the Manim CE API is the right call too. Models already know it, and you don't have to learn a new DSL.

Some caveats. The benchmark numbers come from the project and were run on one machine, so check them on your own hardware. "Compatible with Manim CE's API" doesn't tell you how complete that coverage is, and the README has no limitations section to answer it. Test your existing scenes before you commit. Python 3.13+ is a fairly recent floor. On licensing, the ManimGX code itself is MIT, but the native wheels ship x264, which makes them GPL-3.0-or-later as a whole. That matters if you plan to redistribute.

Who should care: anyone building agents that make math or technical explainer videos, and Manim users who are tired of waiting on renders.