English | δΈζθ―΄ζ | EspaΓ±ol | Deutsch | νκ΅μ΄
π¨ GPT-4o powered generative videos. Concept. β‘οΈ Join our Discord server here!
Generative Manim (GM) is a suite of tools that allows you to create videos with Manim using LLMs (Large Language Models) like GPT-4 or Claude. The idea is to enable anyone to create wonderful animations from text β¨.
It began as a prototype of a web app that uses GPT-4 to generate videos with Manim. The idea behind this project is taking advantage of the power of LLMs in programming, the understanding of human language and the animation capabilities of Manim to generate a tool that could be used by anyone to create videos. Regardless of their programming or video editing skills.
- ποΈ Generative Manim Demo: Check out the demo of Generative Manim!
- π¬ Generative Manim API: Build over the Animation Processing Interface, or API.
- βοΈ Cloud Deployment Guide: Deploy the API on Render or another Docker-based cloud platform.
- π§βπ» Generative Manim Developers: Join our Discord server, learn new things, share your creations and more!
- π Generative Manim Streamlit (Legacy): First LLM exploration of LLMs and Animation.
- 𧩠Manim Awesome Skills: A curated collection of Manim skills and snippets.
Animo is a lightweight desktop application that turns plain English descriptions into animated videos using AI and Manim. It is the natural evolution of Generative Manim, taking everything built in this repository and packaging it into a polished, self-contained desktop experience.
Key improvements over Generative Manim:
- Desktop-first, local rendering: animations render on your machine with no cloud uploads, giving you full privacy and control over your work.
- Unified workflow: everything lives in one app with no need to juggle a web UI, a separate API, and a render environment.
- Zero extra cost: Animo uses your existing Claude or ChatGPT subscription with no additional fees on top of what you already pay.
- Code ownership out of the box: generated Manim code is yours to export, version-control, and reuse without any platform lock-in.
- Lower barrier to entry: no Manim knowledge required and no server setup or Docker container to run.
Models are the core of Generative Manim. A model is a way to convert text to code, that can later be rendered in a video.
| Name | Description | Engine | Phase |
|---|---|---|---|
| GM GPT-5.6 Sol | OpenAI's frontier model for complex professional and coding work | gpt-5.6-sol | β |
| GM GPT-5.6 Terra | Default OpenAI model, balances intelligence and cost | gpt-5.6-terra | β |
| GM GPT-5.6 Luna | Fastest, most cost-efficient OpenAI tier | gpt-5.6-luna | β |
| GM GPT-4o | Latest GPT model from OpenAI powered by a custom System Prompt | GPT-4o | β |
| GM o1-mini | Compact OpenAI reasoning model | o1-mini | β |
| GM GPT-3.5 Fine Tuned | First Fine-tuned model of GPT-3.5 | GPT-3.5 | β |
| GM GPT-3.5 Physics Fine Tuned | Fine-tuned GPT-3.5 model trained to generate Physics animations | GPT-3.5 | β |
| GM Claude Sonnet | Claude Sonnet 3 model from Sonnet adapted with our custom System Prompt | claude-3-sonnet-20240229 | β |
| GM Claude Sonnet 3.5 | Claude Sonnet 3.5 model from Sonnet adapted with our custom System Prompt | claude-3-5-sonnet-20241022 | β |
| GM Claude Sonnet 4.6 | Claude Sonnet 4.6 adapted with our custom System Prompt | claude-sonnet-4-6 | β |
| GM Claude Sonnet 5 | Default Anthropic model, best combination of speed and intelligence | claude-sonnet-5 | β |
| GM Claude Opus 4.7 | Claude Opus 4.7 for highest-quality Anthropic generation | claude-opus-4-7 | β |
| GM Claude Opus 4.8 | Highest capability, state of the art on long-horizon agentic work | claude-opus-4-8 | β |
| GM Claude Haiku 4.5 | Claude Haiku 4.5 for fast, lightweight Anthropic generation | claude-haiku-4-5-20251001 | β |
| GM Claude Fable 5 | Most capable Anthropic model, for the most demanding reasoning | claude-fable-5 | β |
| GM Featherless Open Models | OpenAI-compatible access to hosted open-weight models via Featherless | Qwen, DeepSeek, CodeLlama, etc. | β |
| GM LiteLLM Passthrough | Accepts any model string supported by LiteLLM (e.g. openai/gpt-4o) |
litellm | β |
| GM Gemini 2.5 Flash | Google's Gemini 2.5 Flash accessed via google-genai SDK | gemini-2.5-flash | β |
| GM Gemini 2.5 Pro | Google's Gemini 2.5 Pro accessed via google-genai SDK | gemini-2.5-pro | β |
| GM Gemini 3 Flash | Google's Gemini 3 Flash preview accessed via google-genai SDK | gemini-3-flash-preview | β |
| GM Kimi K3 | Moonshot AI's flagship model, 1M-token context with always-on thinking mode | kimi-k3 | β |
| GM Kimi K2.7 Code | Moonshot AI's coding-specialized agentic model | kimi-k2.7-code | β |
| GM Qwen 2.5 Coder 7B | Open-source model fine-tuned with SFT + DPO + GRPO pipeline | Qwen2.5-Coder-7B-Instruct | π§ |
| GM DeepSeek Coder V2 Lite | Open-source model fine-tuned with SFT + DPO + GRPO pipeline | DeepSeek-Coder-V2-Lite | π§ |
| GM CodeLlama 7B | Open-source model fine-tuned with SFT + DPO + GRPO pipeline | CodeLlama-7b-Instruct | π§ |
If you want to suggest a new model, please open an issue in the repository or talk with us in our Discord server.
We're training open-source models to generate Manim code using a 3-stage pipeline that distills from GPT-4o:
- SFT (Supervised Fine-Tuning): Train on 5,000+ validated promptβcode pairs
- DPO (Direct Preference Optimization): Learn from render success/failure pairs
- GRPO (Group Relative Policy Optimization): RL with the Manim renderer as a deterministic reward signal
The key insight: Manim is a deterministic verifier: code either renders or crashes. This replaces the need for a reward model, similar to how DeepSeek-R1 uses math answer checkers.
Base models: Qwen 2.5 Coder 7B, DeepSeek Coder V2 Lite, CodeLlama 7B. All use QLoRA (4-bit) to fit on free Kaggle T4 GPUs.
Generative Manim now includes an executable benchmark MVP for expert Manim code generation under training/benchmarks.
The benchmark is built around the right primitives for programming evaluation:
- a frozen task suite
- render-based scoring
- Manim-specific structural checks
- pass@k for stochastic code generation
- reproducible JSONL and JSON reports
Start here:
cd training
python -m benchmarks.run export \
--suite benchmarks/tasks/core_v1.jsonl \
--output ./outputs/benchmarks/core_v1_prompts.jsonlThen use the generated prompt file with python -m eval.generate_responses ..., or run the full flow with:
bash ./scripts/run_benchmark.sh qwen2.5-coder-7b ./outputs/grpo/qwen2.5-coder-7b benchmarks/tasks/core_v1.jsonl grpo 5 0.8 1,5See training/benchmarks/README.md for the benchmark design and workflow.
Once you have multiple benchmark runs, compare them with:
cd training
python -m benchmarks.compare --results-dir ./outputs/benchmarks --suite core_v1Or run a whole benchmark matrix from a manifest:
cd training
python -m benchmarks.matrix --manifest benchmarks/manifests/open_source_core_v1.json --dry-runYou can also benchmark hosted open-weight models through Featherless:
export FEATHERLESS_API_KEY="your-featherless-key"
cd training
python -m benchmarks.matrix --manifest benchmarks/manifests/featherless_core_v1.json --only qwen2.5-coder-7b-instruct-featherlessSee docs/featherless.md for API usage, smoke tests, and the full Featherless benchmark workflow.
Generative Manim is currently sponsored by The Astronomical Software Company.
Thank you to everyone who has contributed to Generative Manim!
| Contributor | Contribution |
|---|---|
| @abdullahsohaill | Add Gemini 2.5 Flash and Gemini 3 Flash support via google-genai |
| @tranquac | Fix command injection vulnerability in ffmpeg video export |
| @Wing900 | Add Chinese README translation |
| @Puiching-Memory | Formatter, config, and tooling improvements |
| @RheagalFire | Add LiteLLM as AI gateway engine and streaming/edge case tests |
Generative Manim is an open source project.
If you want to be the author of a new feature, fix a bug or contribute with something new.
Fork the repository and make changes as you like. Pull requests are warmly welcome. Remember you can also join our Discord server to discuss new features, bugs or any other topic.