MiniMax Releases M2.5 Mixture-of-Experts Model

Chinese AI company MiniMax has released MiniMax-M2.5, its first contribution to the open-weight model ecosystem. The new text-generation model employs a Mixture-of-Experts (MoE) architecture, a design known for computational efficiency by activating only specialized parts of the network for any given task.

The most notable technical feature of M2.5 is its use of FP8 weights. By quantizing the model's parameters to an 8-bit floating-point format, MiniMax has significantly reduced the model's memory footprint and potentially increased inference speed compared to models using standard 16-bit formats. This makes the model more accessible to developers and researchers working with consumer-grade hardware.

While the model is publicly available, its utility is limited by its license. MiniMax has released M2.5 under a custom license that prohibits commercial use, positioning it primarily as a resource for academic research and non-commercial experimentation. The full model card and license details are available on its Hugging Face repository.

MiniMax-M2.5's release signals another major AI lab entering the open-weight field, contributing a powerful and efficient architecture. However, its restrictive license highlights a growing trend of "look but don't touch" releases, which allow for research and community evaluation while keeping commercial applications off-limits.