Google DeepMind has opened a public repository for AlphaEvolve, its system that has an AI write programs and keeps the ones that score best, and the repo includes both a working implementation and the files behind its claimed math records. The headline item is a lowered ceiling for the matrix multiplication exponent, a number that measures how fast computers can multiply large grids of numbers in principle. Anyone can now rerun the evidence instead of trusting a press release.

The GitHub page, listed under Google DeepMind (GitHub), is organized as four folders. The first, /src, holds what the authors call a minimal implementation of AlphaEvolve that evaluates candidates locally. The others carry results: /omega for the matrix multiplication bound, /maths for a list of puzzles where the system beat the previous best, and /matmul_tensor_decompositions for the matrix-size results found by its gradient-based method.

The omega folder is the most concrete. Inside it, the evolved optimisation program builds a differentiable model in JAX and Haiku (Google’s numerical Python libraries), then runs gradient descent for 32,000 steps across 32 random starting points. A separate verifier checks the answer with ordinary floating-point arithmetic, which is quick but approximate. For a rigorous proof, a Colab notebook redoes the check in rational arithmetic, and the best bound needs a file called clean_level4.npz.xz that readers must download and upload themselves.

That two-tier setup is the interesting design choice. Fast and loose arithmetic does the searching, and exact arithmetic does the certifying. It means a skeptical mathematician can verify a result without trusting the model that found it, which is exactly what a field wary of AI claims requires.

The maths folder covers more than a dozen problems, among them circle packing, kissing numbers, hexagon packing, cap sets, and the Shannon capacity of cycle graphs. Each has starting programs, the discovered construction, and a Colab notebook that checks it. The repository’s own wording says these are problems where AlphaEvolve outperformed the state of the art, so the verification notebooks are the part worth running rather than the claim.

Running the engine yourself takes more setup. The /src code needs Docker for evaluating candidate programs, Python 3.11, and a Gemini API key, and the example launch command targets the circle packing problem. The model doing the writing is Gemini, so the open code still routes through Google’s API. Teams hoping to run the loop on a local model should not assume the repo supports that.

The README is frank about risk. It warns that AlphaEvolve executes code written by large language models, calls that code unpredictable, and says it should be treated as untrusted. Docker is the only isolation measure, and DeepMind says outright that sandboxing cannot eliminate every chance of unintended behavior or misuse of resources. Google DeepMind also disclaims liability for any damage from running it.

Licensing is permissive. Software falls under Apache 2.0 and other materials under Creative Commons Attribution 4.0, with a 2026 Google copyright line, though a closing disclaimer says the project is no official Google product.

The release is unfinished in a visible way. The paper link and the citation block both still read as TODO notes, so the write-up that is supposed to explain the results is not yet attached to the code. The README gives no timing or cost figures for a run and no independent reproduction.

For anyone who maintains optimisation or scheduling code, the useful experiment is small: point the circle packing example at a problem you already have a scoring function for, and see whether the loop beats your hand-tuned baseline. If it does, the scoring function was the hard part all along.

Reported by Google DeepMind (GitHub) in the AlphaEvolve repository README; the page carries no publication date.