Alpha-Evolve

5/23/2025

A living index of every significant open-source project that reenacts or extends Google DeepMind’s AlphaEvolve agent—complete with plain-English explanations, install notes, and FAQs so any developer can start experimenting.

Updated: 23 May 2025

Why this page?

After DeepMind published AlphaEvolve, developers began releasing their own code so the community could tinker, verify results, and adapt the method to new problems. Below you’ll find every notable repository in one place, plus “at-a-glance” notes on what each project actually does.

Repository spotlights

1. codelion/openevolve

  • Github Repo Link: codelion/openevolve

  • What it is: A “keep-it-simple” port that strips AlphaEvolve down to the essentials—prompt sampling, LLM ensemble, and evolutionary selection. Ideal for readers who want to read every line without sifting through a huge code base.

  • Best for: Teaching, rapid experiments, or building your own lightweight forks.

  • Getting started:

    git clone https://github.com/codelion/openevolve
cd openevolve
pip install -e .
python openevolve-run.py examples/function_minimization/config.yaml

2. puripuriprince/xEvolve

  • Github Repo Link: puripuriprince/xEvolve

  • What it is: A lean replication that emphasises ​throughput​—the author plans to max out concurrency with asyncio. A simple database.json stores ideas so the agent can revisit and mutate them.

  • Best for: Hackers who want to test different evaluation functions quickly or replace the storage layer.

  • Quick run:

    git clone https://github.com/puripuriprince/xEvolve
cd xEvolve
pip install -r requirements.txt
python main.py

  • Roadmap highlights: Better DB backend, ensemble prompting, coverage of “economics-math” and compression tasks.

3. shyamsaktawat/OpenAlpha_Evolve

  • Github Repo Link: shyamsaktawat/OpenAlpha_Evolve

  • What it is: The most mature community project so far—an agent-oriented architecture that mirrors the DeepMind paper almost 1-to-1. Ships with monitoring, RL fine-tuning hooks, and a slick README diagram.

  • Best for: Researchers who want a ready-made sandbox to reproduce the 4×4 matrix-multiplication breakthrough or extend to new algorithm classes.

  • Install & run:

    git clone https://github.com/shyamsaktawat/OpenAlpha_Evolve
cd OpenAlpha_Evolve
pip install -r requirements.txt
python main.py examples/matrix_mul_task.yaml

  • Notable extras:

    • MIT-licensed, so commercial forks welcome.
    • Optional Gemini 2.0 Flash preset configs out of the box.

Frequently-asked questions

> Do any of these repos include the official DeepMind code? No. DeepMind has not released its proprietary training scripts. All projects here are community re-implementations that follow the paper’s public methodology.

> Which repo should I start with as a beginner? Try codelion/openevolve first—it’s under 1 kLOC and easy to read. Once you grasp the loop, move to OpenAlpha_Evolve for larger-scale experiments.

> Can I use GPT-4-o or other OpenAI models instead of Gemini? All three projects rely on generic OpenAI-style SDK calls, so swapping keys/models is usually one command-line flag or config entry.

> How often is this page updated? Weekly during the post-paper “gold-rush” period; afterwards, monthly. Feel free to open a PR or email tips to have a repo listed.

Disclaimer

AlphaEvolve™ is a research project by Google DeepMind. alpha-evolve.com is an independent publication and is not affiliated with or endorsed by Google or DeepMind.

Last refreshed: 23 May 2025

Back to articles