MiniMax-M2.5 NVFP4 Experts-Only PTQ with FP8 KV Cache
This checkpoint was produced using NVIDIA ModelOpt PTQ from a MiniMax-M2.5 source checkpoint.
Quantization summary
- Quantization format: NVFP4
- Scope: MoE experts only
- KV cache: FP8
- Group size: 16
- Calibration algorithm: max
- Calibration size: 512 per default dataset group, 1024 calibration batches total in this run
- Quantized layers: 47,616
- Excluded layers: 311
- Unexpected layers: 0
What is quantized
This checkpoint quantizes only the MiniMax MoE expert projections:
block_sparse_moe.experts.*.w1block_sparse_moe.experts.*.w2block_sparse_moe.experts.*.w3
Both expert weight and input quantizers are NVFP4.
What is not quantized
The following components remain excluded:
- Attention projections:
q_proj, k_proj, v_proj, o_proj
- MoE gate / router path
e_score_correction_biaslm_head
Validation
The checkpoint was structurally validated against the expected MiniMax-M2.5 MoE layout:
- Expected quantized layers:
62 layers × 256 experts × 3 projections = 47,616
- Actual quantized layers:
47,616
- Unexpected layers:
0
The quantized layer set and hf_quant_config.json quantization block match NVIDIA's MiniMax-M2.5-NVFP4 reference metadata. The only observed metadata difference is the local ModelOpt producer version stamp.
Notes
This checkpoint has been structurally validated, but model accuracy and inference quality evaluation have not yet been completed.
