technology
AlphaEvolve
Also known as AlphaEvolve on Google Cloud
AlphaEvolve is a coding agent from Google DeepMind, announced on 14 May 2025. It uses Gemini models to propose programs and automated evaluators to score them.[1][2] Google DeepMind says AlphaEvolve has recovered 0.7% of Google's worldwide compute and sped up a key Gemini training kernel by 23%. It says that, applied to more than 50 open math problems, AlphaEvolve improved on the best known solutions in about 20% of cases.[3][4][5] Google Cloud says it became generally available to Cloud customers in July 2026.[6]
Key facts
AlphaEvolve is an AI agent from Google DeepMind that writes and improves algorithms. Google DeepMind announced it on 14 May 2025.[1]
How it works
AlphaEvolve uses an ensemble of Gemini models. Gemini Flash widens the range of ideas explored, and Gemini Pro adds depth.[1] The models propose programs, and AlphaEvolve verifies, runs and scores each one with automated evaluation metrics.[2] The white paper, posted to arXiv on 16 June 2025, describes a pipeline in which language models change an algorithm’s code directly and evaluators give feedback.[7] Google DeepMind says the agent goes beyond discovering single functions to evolving entire codebases.[8]
Because every candidate has to be scored automatically, the approach suits problems whose solution can be written as an algorithm and checked by a program.[9]
Results in computing
Google DeepMind says a scheduling heuristic found by AlphaEvolve had been in production at Google for over a year by May 2025. It says the heuristic recovers on average 0.7% of Google’s worldwide compute resources.[3] It says an AlphaEvolve rewrite of an arithmetic circuit was integrated into an upcoming Tensor Processing Unit.[10] The agent sped up a key kernel in Gemini’s architecture by 23%, which cut Gemini’s training time by 1%.[4] Google DeepMind also reports up to a 32.5% speedup for a FlashAttention kernel implementation.[11]
Results in mathematics
Google DeepMind says AlphaEvolve found a way to multiply 4x4 complex-valued matrices with 48 scalar multiplications. That improves on Strassen’s 1969 algorithm in that setting.[12] Across more than 50 open problems it rediscovered the best known solutions in roughly 75% of cases and improved on them in about 20%.[5] One improvement was a configuration of 593 spheres that sets a new lower bound for the kissing number problem in 11 dimensions.[13]
A November 2025 paper by Google DeepMind researchers and academic mathematicians applied the agent to 67 problems in analysis, combinatorics, geometry and number theory. It rediscovered the best known solutions in most cases and improved on them in several.[14] The authors also combined AlphaEvolve with Deep Think and AlphaProof to generate proofs automatically.[15] For the wider field, see ai-for-mathematics.
Access
Google released AlphaEvolve to Google Cloud customers in private preview on 10 December 2025.[16] A user supplies a problem specification, evaluation logic and a working seed program. Gemini then generates mutated versions of the code, which are tested against that evaluator.[17] Google Cloud says the service moved to general availability in July 2026.[6]
Read next
AI for mathematics: olympiad proofs, Lean and new results
How AI systems went from olympiad silver to gold, found new constructions with FunSearch and AlphaEvolve, and where research maths still beats them.
Questions readers ask
What is AlphaEvolve?
A coding agent from Google DeepMind that uses Gemini models to propose programs, scores them with automated evaluators and evolves the best ones to improve algorithms.[1][2][7]
What has AlphaEvolve improved at Google?
Google DeepMind says a scheduling heuristic it found recovers on average 0.7% of Google's worldwide compute, a Gemini training kernel ran 23% faster, cutting training time by 1%, and a circuit rewrite went into an upcoming TPU.[3][4][10]
Did AlphaEvolve solve open math problems?
Google DeepMind says that across more than 50 open problems it rediscovered the best known solutions in about 75% of cases and improved on them in about 20%, including a new lower bound for the kissing number in 11 dimensions.[5][13]
Sources
· 17 claims from 4 publishers
Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.
- [1]
Google DeepMind announced AlphaEvolve on 14 May 2025 as a coding agent that uses an ensemble of Gemini models, Gemini Flash for breadth and Gemini Pro for depth, to propose programs. confirmedas of 2025-05-14
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms · Google DeepMind · Primary source · 2025-05-14 · Blog post dated May 14, 2025 (retrieved 2026-11-14)
- [2]
AlphaEvolve verifies, runs and scores the programs it proposes with automated evaluation metrics, which suits domains where progress can be measured clearly. confirmedas of 2025-05-14
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms · Google DeepMind · Primary source · 2025-05-14 (retrieved 2026-11-14)
- [3]
Google DeepMind says an AlphaEvolve scheduling heuristic, in production for over a year by May 2025, recovers on average 0.7% of Google's worldwide compute resources. confirmedas of 2025-05-14
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms · Google DeepMind · Primary source · 2025-05-14 (retrieved 2026-11-14)
- [4]
Google DeepMind says AlphaEvolve sped up a key matrix-multiplication kernel in Gemini's architecture by 23%, cutting Gemini's training time by 1%. confirmedas of 2025-05-14
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms · Google DeepMind · Primary source · 2025-05-14 (retrieved 2026-11-14)
- [5]
Google DeepMind reported in May 2025 that its AlphaEvolve coding agent, applied to more than 50 open problems, rediscovered the best known solutions in roughly 75% of cases and improved on them in 20%. confirmedas of 2025-05-14
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms · Google DeepMind · Primary source · 2025-05-14 (retrieved 2026-11-14)
- [6]
Google Cloud says AlphaEvolve moved from private preview to general availability in July 2026. confirmedas of 2026-11-14
- AlphaEvolve on Google Cloud: AI for agentic discovery and optimization · Google Cloud · Primary source · 2025-12-10 (retrieved 2026-11-14)
- [7]
Google DeepMind's AlphaEvolve white paper, posted to arXiv on 16 June 2025, describes an evolutionary coding agent in which language models change an algorithm's code directly and evaluators give feedback. confirmedas of 2025-06-16
- AlphaEvolve: A coding agent for scientific and algorithmic discovery · arXiv (Google DeepMind) · Primary source · 2025-06-16 (retrieved 2026-11-14)
- [8]
Google DeepMind describes AlphaEvolve as going beyond discovering single functions to evolving entire codebases. confirmedas of 2025-05-14
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms · Google DeepMind · Primary source · 2025-05-14 (retrieved 2026-11-14)
- [9]
Google DeepMind says AlphaEvolve can be applied to any problem whose solution can be written as an algorithm and checked automatically. confirmedas of 2025-05-14
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms · Google DeepMind · Primary source · 2025-05-14 (retrieved 2026-11-14)
- [10]
Google DeepMind says an AlphaEvolve rewrite of an arithmetic circuit was integrated into an upcoming Tensor Processing Unit. confirmedas of 2025-05-14
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms · Google DeepMind · Primary source · 2025-05-14 (retrieved 2026-11-14)
- [11]
Google DeepMind says AlphaEvolve achieved up to a 32.5% speedup for a FlashAttention kernel implementation. confirmedas of 2025-05-14
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms · Google DeepMind · Primary source · 2025-05-14 (retrieved 2026-11-14)
- [12]
Google DeepMind says AlphaEvolve found a way to multiply 4x4 complex-valued matrices using 48 scalar multiplications, improving on Strassen's 1969 algorithm in that setting. confirmedas of 2025-05-14
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms · Google DeepMind · Primary source · 2025-05-14 (retrieved 2026-11-14)
- [13]
Google DeepMind says AlphaEvolve found a configuration of 593 outer spheres, establishing a new lower bound for the kissing number problem in 11 dimensions. confirmedas of 2025-05-14
- AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms · Google DeepMind · Primary source · 2025-05-14 (retrieved 2026-11-14)
- [14]
A November 2025 paper by Google DeepMind researchers and academic mathematicians applied AlphaEvolve to 67 problems in analysis, combinatorics, geometry and number theory; it rediscovered the best known solutions in most cases and improved on them in several. confirmedas of 2025-11-03
- Mathematical exploration and discovery at scale · arXiv · Primary source · 2025-11-03 (retrieved 2026-11-14)
- [15]
The same paper combined AlphaEvolve with Deep Think and AlphaProof to add automated proof generation. confirmedas of 2025-11-03
- Mathematical exploration and discovery at scale · arXiv · Primary source · 2025-11-03 (retrieved 2026-11-14)
- [16]
Google released AlphaEvolve to Google Cloud customers in private preview on 10 December 2025. confirmedas of 2025-12-10
- AlphaEvolve on Google Cloud: AI for agentic discovery and optimization · Google Cloud · Primary source · 2025-12-10 (retrieved 2026-11-14)
- [17]
On Google Cloud, a user supplies a problem specification, evaluation logic and a working seed program, and AlphaEvolve's Gemini models generate mutated versions of the code that are tested against that evaluator. confirmedas of 2025-12-10
- AlphaEvolve on Google Cloud: AI for agentic discovery and optimization · Google Cloud · Primary source · 2025-12-10 (retrieved 2026-11-14)
- AlphaEvolve on Google Cloud: AI for agentic discovery and optimization · Google Cloud · Primary source · 2025-12-10 (retrieved 2026-11-14)
Revision history (1)
- Page created.
Created Nov 14, 2026. Last reviewed by an editor on Nov 15, 2026. Next scheduled review: Nov 14, 2027.
Cite this page
"AlphaEvolve." ContentLora, updated Nov 14, 2026. https://contentlora.com/wiki/alphaevolve
Spotted an error? Suggest a correction or emailcorrections@contentlora.com.