Model Releases
MiniMax Releases M3.1 Flash 428B MoE Language Model
MiniMax officially launches M3.1 Flash, a 428B MoE model optimized for agentic workflows, coding, and computer use with 3x faster inference.
October 19, 2026
Model ReleaseMiniMax M3.1 Flash
Introduction
The AI industry has long awaited the official identity of 'Space Bunny,' the mysterious anonymous model that recently made waves by processing over 200 trillion tokens in a massive, high-volume stress test. Today, MiniMax has officially pulled back the curtain, announcing the release of MiniMax M3.1 Flash. This is not just a rebranding; it is the culmination of massive-scale training and architectural refinement designed specifically for the era of autonomous agents.
As we move deeper into 2026, the bottleneck for AI development has shifted from raw reasoning to the cost and speed of agentic execution. M3.1 Flash addresses this head-on, offering a generational leap in capabilities for tasks requiring computer use, cybersecurity, and complex software engineering. By prioritizing high-speed online inference and massive efficiency, MiniMax is positioning this model as the backbone for the next generation of autonomous SaaS workflows.
- Official release of the 'Space Bunny' lineage
- Optimized for agentic, coding, and computer-use tasks
- 3x faster online inference compared to M3 predecessors
Key Features & Architecture
At the core of MiniMax M3.1 Flash is a sophisticated 428B Mixture-of-Experts (MoE) architecture. Unlike dense models that require massive compute for every token, this MoE approach allows the model to activate only a fraction of its parameters for any given task, significantly reducing latency and operational costs without sacrificing the depth of knowledge required for complex reasoning.
The model is specifically tuned for multimodal interaction and high-frequency tool calling. While the architecture is optimized for speed, it maintains a massive parameter count that allows it to handle intricate instructions and long-context dependencies. This makes it uniquely suited for 'Computer Use' scenarios where the model must interpret UI elements and execute multi-step actions in real-time environments.
- 428B Parameter Mixture-of-Experts (MoE) architecture
- Optimized for low-latency online inference
- Advanced multimodal and tool-calling capabilities
- High-speed execution for agentic loops
Performance & Benchmarks
M3.1 Flash has demonstrated industry-leading performance across several critical benchmarks for automation and software engineering. On AutomationBench-AA, the model ranks among the top performers, proving its ability to complete end-to-end enterprise workflows across diverse sectors including sales, marketing, finance, and HR. It doesn't just predict text; it navigates SaaS systems to achieve business objectives.