Compatible MiniMax Models

You can import large language models from Hugging Face and OCI Object Storage buckets into OCI Generative AI, create endpoints for those models, and use them in the Generative AI service.

MiniMax M3

The MiniMax-M3-MXFP8 model is the MXFP8 quantized variant of MiniMax M3, a native multimodal model with a one million token context. The model has about 428 billion total parameters with about 23 billion activated parameters and uses MiniMax Sparse Attention (MSA) for efficient long-context processing. This model is optimized for coding, long-horizon agentic workflows, and collaborative productivity tasks.

Compatible MiniMax M3 Models
Hugging Face Model ID Model Capability Recommended Dedicated AI Cluster Unit Shapes
MiniMaxAI/MiniMax-M3-MXFP8 TEXT_TO_TEXT
  • H200_X8
  • B200_X8

MiniMax M2

The MiniMax M2 text-to-text models are optimized for coding, complex reasoning, and agentic workflows such as tool use, search, and productivity tasks. MiniMax-M2 is a Mixture-of-Experts (MoE) model designed for efficient coding and agentic performance, and later MiniMax-M2 models extend this focus to more advanced software engineering and professional-work tasks. For more details, see MiniMax in the Hugging Face documentation.

Compatible MiniMax M2 Models
Hugging Face Model ID Model Capability Recommended Dedicated AI Cluster Unit Shapes
MiniMaxAI/MiniMax-M2.7 TEXT_TO_TEXT
  • H100_X8
  • H200_X8
MiniMaxAI/MiniMax-M2.5 TEXT_TO_TEXT
  • H100_X8
  • H200_X8
MiniMaxAI/MiniMax-M2 TEXT_TO_TEXT
  • H100_X8
  • H200_X8
Important

  • For imported models, you can use the native context length specified by the model provider. However, the effective maximum context length is limited by the underlying hardware setup that you select for the hosting dedicated AI clusters in OCI Generative AI. To take full advantage of a model's native context length, you might need to provision more hardware resources.
  • Use the fine-tuned models only if they match the compatible base model's transformer version and have a parameter count within ±10% of the original.
  • For available hardware and steps on how to deploy the imported models, see Managing Imported Models.