Gemini

Type: Product (Large Language Model Family)

Gemini is Google DeepMind's family of large language models. In this knowledge base, it appears primarily as the engine powering AlphaEvolve's evolutionary code generation.

The AlphaEvolve system uses a Flash/Pro ensemble strategy: Gemini Flash maximizes the breadth of ideas explored (generating many candidate solutions quickly), while Gemini Pro provides critical depth with insightful suggestions (evaluating and refining the most promising candidates).

The 2026 upgrade to Gemini 2.5 Pro enabled semantic evolution — the model can now reason about code logic and control flows rather than just adjusting parameter values. This qualitative leap allows AlphaEvolve to discover fundamentally different algorithmic approaches.

Key Contributions

• Flash/Pro ensemble for evolutionary search: Breadth (Flash) + depth (Pro) strategy in AlphaEvolve (AlphaEvolve)
• Gemini 2.5 Pro semantic evolution: Logic/control flow rewriting, not just hyperparameter tuning (AlphaEvolve)
• Self-improving loop: AlphaEvolve optimizes Gemini's own training kernels (23% speedup, 1% training time reduction) (AlphaEvolve)

Mentioned In

• Evolutionary Algorithm Discovery — Powers the evolutionary search
• LLM Tool Use — Code generation as tool use

Related Entities

• Google DeepMind — Developer
• AlphaEvolve — Primary application documented in this KB