MiniMax-M2

MiniMax-M2 is a large sparse MoE language model from MiniMaxAI. Megatron Bridge supports MiniMax-M2 and architecture-compatible M2.5/M2.7 checkpoints through the MiniMaxM2Bridge.

Supported Variants

Variant	Hugging Face ID	Notes
MiniMax-M2	MiniMaxAI/MiniMax-M2	456B total, 45.9B active
MiniMax-M2.5	MiniMaxAI/MiniMax-M2.5	Same MiniMaxM2ForCausalLM architecture
MiniMax-M2.7	MiniMaxAI/MiniMax-M2.7	Same MiniMaxM2ForCausalLM architecture

Architecture Notes

• Sparse MoE with 256 experts and top-8 routing.

• Sigmoid router with expert-bias correction.

• FP8 block-wise checkpoint weights are dequantized during import.

• Uses full-dimension QK RMSNorm; the bridge provides custom Q/K norm mappings for tensor-parallel shards.

• MTP modules are not currently mapped.

Examples

For Slurm conversion, inference, hardware requirements, and validated parallelism settings, see the MiniMax-M2 examples README.

Related Implementation

• Bridge implementation: src/megatron/bridge/models/minimax_m2

• Examples: examples/models/minimax/minimax_m2