Enhancing AI Collaboration: Unveiling the Mixture of Agents (MoA) Approach

In the ever-evolving landscape of AI, the newly introduced Mixture of Agents (MoA) represents a significant step forward in leveraging the collective strengths of multiple large language models (LLMs) to achieve enhanced performance and robustness. The core idea behind MoA is to create a synergistic system where multiple LLMs collaborate, each enhancing the overall capabilities of the ensemble through a structured, layered process.

What is Mixture of Agents (MoA)?

MoA operates on a principle termed the “collaborativeness of LLMs,” which suggests that an LLM can produce superior outputs when it processes and refines the responses generated by other models, regardless of the standalone capabilities of these contributing models. This phenomenon is leveraged by categorizing the models into two functional groups: proposers and aggregators.


These are models tasked with generating initial responses. Each proposer may offer unique insights and perspectives, which are crucial for providing a diverse range of potential answers or solutions to a given query.


Aggregators synthesize the responses from one or more proposers, refining them into a singular, high-quality output. This stage is critical for integrating the diverse inputs into a coherent and contextually accurate response.

Implementation and Performance

The reference implementation, dubbed Together MoA, incorporates this architecture using several open-source LLMs in a configuration that spans multiple layers. Initial tests have shown that Together MoA achieves a 65.1% score on AlpacaEval 2.0, which is a notable improvement over the previous leader, GPT-4o, which scored 57.5%.

The Layered Approach

The MoA structure is built with several layers, each comprising a set of proposers followed by an aggregator. This multi-layer approach allows for iterative refinement of responses, where each layer aims to enhance the quality and accuracy of the output from the preceding layer. This structured refinement is pivotal in achieving the high accuracy and robustness of the model.

Results and Evaluation

The efficacy of the Together MoA model has been rigorously tested against established benchmarks such as AlpacaEval 2.0, MT-Bench, and FLASK. These evaluations demonstrate significant improvements across various metrics, including correctness, factuality, and insightfulness. For instance, in the AlpacaEval 2.0, Together MoA outperformed GPT-4o by a substantial margin.


One of the considerations in the development of MoA has been the balance between performance and cost. While high-performance configurations of MoA achieve the best results, there are also “Lite” versions which offer a more cost-effective solution without significantly compromising the output quality. This makes MoA adaptable to different usage scenarios and budget constraints.

The Collaborative Ecosystem

The development of Together MoA has been supported by a collaborative effort among several leading organizations in the AI space. These include contributions from Meta AI, Microsoft, and smaller specialized groups like DataBricks and Mistral AI. Such collaboration underscores the community-driven aspect of AI development, particularly in the open-source domain.

Future Directions

Looking forward, the team behind MoA plans to explore various enhancements, such as reducing the latency in initial response generation and optimizing the model architecture for more reasoning-intensive tasks. These improvements aim to refine MoA’s effectiveness and extend its applicability to more complex AI challenges.


Together MoA exemplifies how collaborative efforts in the AI community can lead to advancements that are not only technically proficient but also broadly accessible and adaptable. By integrating the capabilities of multiple LLMs, MoA sets a new standard for what is achievable in the realm of artificial intelligence.

For a more in-depth exploration of the MoA approach and to view the interactive demo, visit the project’s page at Together MoA.