AlphaEvolve

Type: Product (Evolutionary Coding Agent)

AlphaEvolve is Google DeepMind's evolutionary coding agent for general-purpose algorithm discovery and optimization. It pairs an ensemble of Gemini models (Flash for breadth, Pro for depth) with automated evaluators, using evolutionary frameworks to iteratively improve algorithmic solutions.

Deployed inside Google's production infrastructure for over a year, it has demonstrated both mathematical breakthroughs (improving on Strassen's 57-year-old matrix multiplication algorithm) and practical production impact (recovering 0.7% of Google's global compute, speeding up Gemini's own training).

The 2026 upgrade to Gemini 2.5 Pro introduced semantic evolution — rewriting logic and control flows rather than just tuning parameters. The AlphaEvolve Service API is now available via Google Cloud Early Access, and the open-source OpenEvolve project provides a community implementation.

Key Contributions

• Improved on Strassen's 1969 matrix multiplication algorithm (48 scalar multiplications for 4x4 complex) (AlphaEvolve)
• 0.7% continuous recovery of Google global compute resources (AlphaEvolve)
• 23% kernel speedup → 1% Gemini training time reduction (AlphaEvolve)
• 32.5% FlashAttention kernel speedup (AlphaEvolve)
• Matched SOTA on ~75%, improved ~20% of 50+ open math problems (AlphaEvolve)

Mentioned In

• Evolutionary Algorithm Discovery — Primary instantiation
• Agentic Reasoning — Example of production agentic system
• LLM Tool Use — Code generation as specialized tool use

Related Entities

• Google DeepMind — Developer
• Gemini — Underlying model ensemble