Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It — Olive Song

MiniMax chose to open source their most advanced model, M3, believing in the power of the open-source community. This decision aligns with their mission to make intelligence accessible to everyone and allows developers to contribute feedback and improvements. The company anticipates that releasing the model openly will also enable optimization by partners like Kernels Together AI for faster inference and broader usage.

The partnership between MiniMax and Kernels Together AI began after a meeting at a car event in Las Vegas. A representative from MiniMax expressed excitement about the upcoming M3 model and encouraged Kernels Together AI to serve it, highlighting its potential. This initial conversation led to collaborative efforts on optimizing the model's architecture and inference capabilities.

A key differentiator for MiniMax M3 is its multi-modal capability, allowing it to understand not only text and code but also videos and images. This has enabled innovative applications such as agent development that can navigate computer interfaces and even game creation, representing a significant advancement in model functionality beyond previous iterations like M2.

Olive Song emphasizes that the quality and design of training data are crucial for achieving high performance on specialized benchmarks such as KernelBench and OSWorld. The process involves carefully defining problem environments within the data to ensure the model learns effectively, highlighting a key aspect of post-training strategy.

MiniMax employs an iterative evaluation process when dealing with long-horizon tasks, such as those lasting 12 hours. The model submits multiple iterations during these extended runs, and each submission is evaluated individually to identify potential issues like 'hacking' or unexpected behavior. This allows for validation and testing to ensure genuine performance improvements rather than unintended exploits.

Addressing the challenge of a growing knowledge base (KB) cache with concurrent requests requires building infrastructure akin to a distributed file system or large database. Dan explains that this involves complex considerations like storage location, data retrieval mechanisms, and efficient transfer protocols – concepts often overlooked during undergraduate studies but crucial for scaling AI systems.

Looking ahead three years, Dan anticipates significant improvements in GPU utilization within the AI field. He references SpaceX's current low utilization rate (around 10%) and expresses hope that the industry will be 'embarrassed' by its performance in comparison. This suggests a focus on optimizing training processes to maximize hardware efficiency.

Olive highlights the accelerating pace of AI development, particularly driven by techniques like self-evolution. She notes that models developed even a year ago have already significantly improved development speed and enabled open-weight models to rapidly close the gap with leading labs. This underscores MiniMax's commitment to democratizing access to advanced AI technology.