# Model Mixing
**Model Mixing** refers to strategies and techniques for combining multiple [[concepts/large-language-model]]s (LLMs) or leveraging different models within a single system to optimize for specific tasks, balancing cost, latency, and performance.
## Core Principles
- **[[concepts/specialization|Specialization]] vs. Generalization**: Using smaller, [[concepts/custom-models|specialized models]] for rote tasks (e.g., formatting, parsing) and larger, more capable models for [[concepts/complex-reasoning|complex reasoning]] or creative generation.
- **Ensemble Methods**: Aggregating outputs from multiple models to reduce [[concepts/data-hallucination|hallucination]] rates and improve [[concepts/robustness|robustness]].
- **Routing**: Dynamically directing queries to the most appropriate model based on input complexity or user constraints.
## Critiques and Alternatives
Recent analysis suggests that the efficacy of AI systems may depend less on the raw capability of the underlying LLM and more on the structural design of the surrounding system.
- **The [[concepts/harness|Harness]] Hypothesis**: The "harness" ([[concepts/context-management|context management]], tool usage, error correction [[concepts/loops|loops]]) often outweighs the marginal gains of switching between top-tier LLMs. See [[lab-notes/2026-05-18-Optimizing-AI-Coding-Agents-Harness-Design-Over-LLM-Choi|Optimizing AI Coding Agents: Harness Design Over LLM Choice]] for details on how [[concepts/ai-prompt|prompt architecture]] and agent [[concepts/open-source-philosophy|logic]] can mitigate model limitations.
- **[[concepts/diminishing-returns|Diminishing Returns]]**: Blindly mixing models without optimizing the orchestration layer can lead to increased latency and complexity without proportional quality improvements.
## Related Concepts
- Agent Design
- [[entities/prompt-engineering]]
- LLM [[concepts/benchmark-testing|Benchmarking]]
- [[concepts/system-prompting]]