Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It — Olive Song
Olive Song, RL lead at MiniMax, details the engineering behind MiniMax's open-weight models, focusing on M3's multimodal and agentic capabilities, the necessity of day-zero inference stack readiness, and continuous GPU kernel optimization. She discusses multimodal training challenges, long-horizon task evaluation, and expresses optimism for open models rapidly closing the gap with frontier labs.
Olive Song, Reinforcement Learning lead at MiniMax, provides an in-depth look into the company’s open-weight models and the sophisticated infrastructure that supports them. A foundational belief in open source drives MiniMax’s strategy, aiming to democratize intelligence and empower developers to contribute and optimize, thereby enhancing the models and accelerating inference.
Key Capabilities and Applications of MiniMax M3
MiniMax M3 stands out from its M2 predecessors as a multimodal model, proficient in understanding text, writing code, and processing videos and images. This multimodal capability unlocks several advanced applications:
- Agentic Multimodal Agents: The model effectively performs agentic tasks by integrating multiple data modalities.
- Computer Use: M3 can navigate and interact with computer environments, enabling complex creations through available tools, as demonstrated in benchmarks like OS World.
- Game Development: A notable, and perhaps unexpected, application is the model’s ability to assist in developing games.
Training Multimodal Models and Long-Horizon Tasks
Developing such capabilities requires meticulously designed data and environments. For reinforcement learning (RL) tasks, it is crucial to formulate problems, define rewards, and adapt RL algorithms for efficient training. Song highlights a critical challenge in multimodal training: the tendency for text and vision modalities to “collapse after training a little bit” if not trained conjointly. MiniMax M3 was trained multimodally from scratch, ensuring that “the text tokens would attend to the visual tokens,” fostering intrinsic understanding and integration of modalities.
The model’s capacity for long-horizon tasks, such as replicating a 12-hour run to reproduce an ICLR paper, underscores its advanced reasoning. This involves iterative evaluation, where the model submits multiple solutions, and internal evaluations rigorously test for genuine performance improvements, guarding against mere “hacking” or overfitting. MiniMax actively employs its models for “self-evolution,” using them to accelerate internal development.
The Engineering Imperative: Inference Stack and Optimization
Delivering models like M3 to users necessitates an exceptionally optimized inference stack, prepared for “day zero” launch:
- Early Access and Kernel Development: Partners receive early architectural details (e.g., MiniMax’s sparse attention, MOE choices, quantization specifics) to proactively develop or adapt GPU kernels.
- Rigorous Benchmarking: Comprehensive benchmarking, including Together AI’s parallel kernel bench, is vital. This benchmark identifies unsolved problems, and overfitting to it is encouraged as it directly leads to performance improvements in real-world use cases.
- Continuous Optimization: From the initial launch, a prioritized list of optimizations is tackled, covering aspects such as KV cache handling, attention kernels, and quantization, ensuring that performance progressively improves “week over week, month over month.”
- Adapting to Workload Shifts: The evolving landscape from primarily chat-based interactions to agentic workflows (e.g., coding agents processing large codebases, multi-turn tool calls) profoundly influences inference stack design. This demands distinct optimization strategies for KV cache, prompting, routing, and kernel challenges compared to traditional chat scenarios.
- Holistic Optimization: Faced with myriad optimization opportunities, the approach is to tackle an exhaustive list. Lessons from optimizing sparse attention mechanisms in one model can often be adapted to others, even with architectural variations.
Future Outlook
Both speakers express significant optimism for the future of AI. Song observes the accelerating pace of development, which enables open-weight models to “catch up with frontier labs” much faster than previously anticipated. Dan anticipates substantial improvements in GPU utilization, moving beyond current inefficiencies, and believes that “open models are closing the gap faster than a year ago would have suggested,” challenging the notion of closed models’ inherent superiority.