The Quiet Revolution of MiniMax: Meet M3.1 Flash and the Secret Behind "Space Bunny"
The world of language models continues to move at a dizzying pace, and MiniMax, one of the most intriguing players in the field, is now landing a significant update with the launch of the M3.1 Flash. If the name "Space Bunny" sounds familiar from discussions in developer communities recently, you are not wrong – the company has revealed that it was an early version of the new model, which operated under a veil of anonymity and processed over 200 trillion tokens before the official disclosure [1, 2, 10].
Why does this matter?
For developers and AI professionals, the M3.1 Flash is not just "another model." It was specifically designed for complex tasks requiring agentic AI capabilities, the use of external tools, code generation, and interaction with computer environments [1, 7, 10]. The combination of improved performance and a dramatic reduction in costs makes it an especially attractive work tool for those seeking efficiency without compromise [2, 7].
Performance, Speed, and Costs
The data presented by MiniMax is impressive: the new model offers inference speed that is 3 times faster compared to the previous generation (M3), while reducing usage costs by about 70% [1, 2, 7].
The new API pricing structure looks like this:
- •Input: $0.15 per million tokens [1, 10].
- •Output: $0.60 per million tokens [1, 10].
- •Cached Input: $0.015 per million tokens – a significant 90% discount compared to standard input [1, 7].
Beyond speed, the model exhibits high capabilities in benchmarks within the fields of programming and automation, ranking high among the leading Chinese models [2, 10]. It supports a context window of up to one million tokens, allowing it to handle long-term projects and complex tasks [1, 7].
Availability and Next Steps
The model is already available via the MiniMax API, the M Plan platform, and the dedicated development tool Qoder [2, 7, 10]. One of the most interesting steps is the company's commitment to releasing the model weights to the public, which opens the door for developers to run the model independently and perform custom optimizations [1, 7].
Summary and Conclusions
The launch of M3.1 Flash highlights a clear trend: the transition from general models to task-specific models that are tailored for working with agents. The fact that MiniMax chose to test the model "in the field" under the name Space Bunny testifies to great confidence in the system's capabilities even before the official launch [1, 2].
For those involved in no-code development or building AI agents, this is significant news. The ability to run complex tasks at a significantly lower cost allows for expanding the boundaries of automation in organizations and personal projects.
Additional sources for review: