πŸ‡ΈπŸ‡¬

MiniMax on Berges AI

The open-weights model from MiniMax, built on linear-attention mixture-of-experts. Currently running in the Berges deep-thinking layer.

MiniMax is an AI lab headquartered in Singapore. The MiniMax model family is known for fast generation and large effective context, built on a linear-attention mixture-of-experts architecture that handles long inputs efficiently.

Berges AI runs MiniMax in the deep-thinking layer alongside Kimi and DeepSeek-Pro. You can pick it explicitly from the sidebar if you want to see how the same prompt plays out across the three.

The weights are open and the architecture is documented publicly.

At a glance

MiniMax specs.

Maker
MiniMax (Singapore)
License
Open weights
Architecture
Linear-attention Mixture-of-Experts
Layer on Berges
Deep-thinking
Current version
M2.7
Best known for
Fast generation, long context
Strengths

What MiniMax is good at.

Fast generation

Linear attention scales better with context length than standard attention. The practical effect is quicker responses on long-form generation.

Long context

MiniMax handles long inputs efficiently, so it does well on tasks where you're feeding the model a lot of material to reason over.

Multilingual

Strong on Chinese-English work and reasonable across other major languages.

Reach for it when

MiniMax is the right pick.

  • β†’ Long-form generation where speed matters
  • β†’ Long-context summarization or reasoning
  • β†’ Chinese-English translation and writing
Questions

Things people ask about MiniMax.

Where is MiniMax based?

MiniMax is headquartered in Singapore. The model weights are open and the company publishes details about its architecture publicly.

Why pick MiniMax over Kimi or DeepSeek?

On Berges AI you don't have to. The cascade handles it. If you do pick manually, MiniMax tends to be faster on long-form generation, while Kimi is stronger on long-context reasoning and DeepSeek on math and code.

Is MiniMax open source?

The weights are open. The training data is generally not, which is typical for the open-weights category. You can audit and run the model, just not retrain it from scratch with the same data.