Google DeepMind has published an official open-source companion repository for AlphaEvolve, its evolutionary coding agent for scientific and algorithmic discovery. The google-deepmind/alphaevolve repository appeared on November 4, 2026 under the Apache-2.0 licence for software and CC-BY for other materials. DeepMind labels the package as not an official Google product.
The /src tree contains a minimal AlphaEvolve implementation meant for local evaluation. Setup instructions ask for Docker because candidate programs generated by a language model run inside a container as an isolation measure, a Gemini API key, and pinned Python dependencies. DeepMind warns that LLM-generated code remains untrusted even when sandboxed and that users accept residual risk.
Alongside the agent loop, the repository ships verification assets for results AlphaEvolve previously reported. The /omega directory holds code and data for a new upper bound on the matrix-multiplication exponent, including a JAX and Haiku optimisation program evolved by AlphaEvolve, a NumPy and SciPy verifier, a Dockerfile, and a Colab notebook that uses rational arithmetic for a stricter proof. Folders under /maths and /matmul_tensor_decompositions provide initial programs, discovered constructions and Colab notebooks for problems where AlphaEvolve beat prior state of the art, from circle packing and kissing numbers to cap sets and matrix-multiplication tensor decompositions.
Earlier public DeepMind artefacts such as alphaevolve_results shared discoveries without the runnable agent. This repository is the first DeepMind GitHub tree that includes a local AlphaEvolve implementation together with those verification notebooks. The README still contains TODO markers for the paper link and citation block, so bibliographic details were incomplete at publication time.
Independent mirrors summarised the README on the same day; substantive journalistic follow-ups were still thin at scout time, so confidence rests mainly on the first-party DeepMind organisation release. For researchers the practical value is reproducibility: a Docker-gated local loop plus checkable maths artefacts, not a managed cloud service. Separate Google Cloud client libraries for a hosted AlphaEvolve offering remain a different codebase and should not be confused with this research dump.
Practically, the release closes a gap that frustrated outsiders after the original AlphaEvolve write-ups: results notebooks existed, but the evolutionary loop did not. Researchers can now inspect how problems are configured, how candidates are evaluated under Docker, and how claimed constructions are checked. That does not turn the dump into a product, and it does not fill in the missing paper URL, yet it is the strongest public starting point DeepMind has offered for hands-on AlphaEvolve work so far.