Google DeepMind Opens AlphaEvolve to Everyone After Doubling Klarna's AI Throughput

AlphaEvolve, Google DeepMind's Gemini-powered algorithm optimization agent, is now generally available on Google Cloud. What started as an internal research tool quietly improving Google's own data centers and chip designs has graduated into a full cloud product, accessible to any organization willing to point it at their hardest engineering problems.

Not your average coding assistant

AlphaEvolve is not a code completion tool. It is an evolutionary coding agent -- a system that autonomously generates, tests, and iteratively refines entire algorithms through a loop of mutation and selection, guided by a scoring function you define. Think of it as a search engine for better code, one that explores a combinatorially vast design space no human team could comb through manually.

AlphaEvolve pairs the creative problem-solving capabilities of Gemini models with automated evaluators that verify answers, and uses an evolutionary framework to improve upon the most promising ideas. It leverages an ensemble of models: Gemini Flash maximizes the breadth of ideas explored, while Gemini Pro provides critical depth with insightful suggestions.

The key mechanic that separates this from a standard LLM coding agent is the evaluator. AlphaEvolve combines creative, server-side LLM exploration with secure client-side code execution to autonomously discover new, optimized solutions that surpass human-designed baselines. Your evaluator runs locally, on your own infrastructure, scoring each candidate the system proposes.

The evolutionary loop, in practice

Getting started requires just two inputs:

1. A seed program: Your existing algorithm, written as code. You mark which sections are open for optimization.
2. An evaluator: A deterministic script that compiles, runs, and scores each candidate solution, returning scalar metrics for AlphaEvolve to maximize.

From there, AlphaEvolve runs an autonomous loop. Gemini Flash handles speed, generating mutated and optimized versions of the code added to the population space, while Gemini Pro provides higher-quality suggestions. Evolution algorithms then select which mutations to combine and further mutate, prioritizing them as the starting point for the next generation.

# Minimal AlphaEvolve interaction pattern
import alphaevolve
client = alphaevolve.Client()
# 1. Register your seed algorithm and evaluator
session = client.create_session(
    seed_program="my_algorithm.py",
    evaluator_endpoint="http://localhost:8080/evaluate"
)
# 2. Sample a candidate from the evolutionary search
candidate = session.sample_candidate()
# 3. Run your evaluator locally, submit the score back
score = my_local_evaluator(candidate.code)
session.submit_score(candidate.id, score)
# Repeat until convergence

A year of production evidence

AlphaEvolve was developed by Google DeepMind and unveiled in May 2025. The GA launch is backed by a substantial track record. Over the past year, Google deployed algorithms discovered by AlphaEvolve across its computing ecosystem, including data centers, hardware, and software. The internal results are striking:

• AlphaEvolve found a better way to schedule tasks in Google's data centers, continuously recovering on average 0.7% of global compute resources.
• It sped up a vital kernel in Gemini's architecture by 23%, leading to a 1% reduction in Gemini's training time.
• In quantum physics, AlphaEvolve's optimizations made it possible to run complex molecular simulations on Google's Willow quantum processor by suggesting quantum circuits with 10x lower error than previous conventionally optimized baselines.
• In grid optimization, it helped increase the ability of a trained Graph Neural Network to find feasible solutions for the AC Optimal Power Flow problem from 14% to over 88%.

There is also a self-referential quality worth noting: AlphaEvolve improved the pipeline used to train Gemini -- the same model used inside AlphaEvolve itself.

Enterprise early adopters are already seeing gains

The GA launch comes with a wall of real customer results across logistics, finance, genomics, and semiconductors. These are not toy benchmarks.

• Klarna: Applied AlphaEvolve to one of their largest ML training pipelines and doubled throughput while improving model quality. Over three weeks, the system explored nearly 6,000 candidate programs, discovering deep architectural rewrites no engineer would have tried.
• Kinaxis: Achieved improvements of more than 22% in key forecasting accuracy metrics while reducing runtime by over 90% on benchmark datasets.
• FM Logistic: Achieved a 10.4% improvement in warehouse routing on top of an already highly optimized baseline, saving over 15,000 km in staff travel.
• Schrödinger: Used AlphaEvolve to quadruple the speed of molecular force field inference, shortening drug discovery R&D cycles from months to days.
• JetBrains: Improved IDE algorithm performance by 15-20% on problems that were previously too time-consuming to optimize manually.
• BASF: Built a working digital twin of their global supply chain network after previous deterministic modeling attempts had all failed.
• PacBio: Used AlphaEvolve to improve DeepConsensus -- a model for correcting DNA sequencing errors -- achieving a 30% reduction in variant detection errors.

The research lineage behind the product

Unlike domain-specific predecessors like AlphaFold or AlphaTensor, AlphaEvolve is designed as a general-purpose system that can operate across a wide array of scientific and engineering tasks by automatically modifying code and optimizing for multiple objectives. Its architecture allows it to evaluate code programmatically, reducing reliance on human input and mitigating risks such as hallucinations common in standard LLM outputs.

The broader research community has taken note. AlphaEvolve's closed-source nature and reliance on proprietary models has spurred open-source alternatives, including OpenEvolve, CodeEvolve, and ShinkaEvolve. The GA launch now gives the enterprise world direct access to the original, backed by Google Cloud's SLA and infrastructure.

What this unlocks -- and what it doesn't

AlphaEvolve can be applied to any problem whose solution can be described as an algorithm and automatically verified. That is a broad mandate, but it comes with a hard constraint: if you cannot write an objective, deterministic scoring function, AlphaEvolve cannot help you. Subjective quality, user experience, and open-ended creativity are out of scope.

The domains where it shines are exactly the ones where optimization has historically been bottlenecked by search-space size:

• GPU kernel and compiler optimization
• ML training pipeline architecture
• Logistics routing and scheduling
• Molecular simulation and drug discovery
• Chip design and hardware arithmetic circuits
• Demand forecasting and supply chain planning

Note that AlphaEvolve does not support FedRAMP or DoD compliance requirements, and access for environments requiring these standards is restricted by default. Organizations in regulated sectors will need to factor that in before committing to production workflows.

The bigger picture

The GA launch signals something larger than a product release. AlphaEvolve is a strong sign that the next enterprise AI frontier is not only better conversational models -- it is AI systems that can discover measurable improvements inside real operating environments. The agent-as-optimizer pattern, where an LLM-powered system is paired with a rigorous evaluator and given autonomy to search, is maturing into a deployable primitive.

As Pushmeet Kohli, Chief Scientist at Google Cloud and VP at Google DeepMind, put it: AlphaEvolve has become a powerful engine for scientific and societal progress -- and now, for the first time, that engine is available to anyone with a Google Cloud account and a hard optimization problem to solve. You can get started through the official documentation and the open example repository.