MiniMax-M2.5 Leaked: 230B Parameters with Only 10B Active
A massive new model architecture leak could change the scaling game...
Deep Dive
A leak from OpenHands reveals MiniMaxAI's upcoming MiniMax-M2.5 model reportedly contains a staggering 230 billion total parameters, but only activates 10 billion for inference. This suggests a novel, highly efficient Mixture of Experts (MoE) architecture. The model details were announced ahead of its expected release on Hugging Face, sparking immediate speculation about its performance and efficiency benchmarks compared to giants like GPT-4 and Claude 3.
Why It Matters
If true, this architecture could deliver top-tier performance at a fraction of the computational cost, making powerful AI more accessible.