num_local_experts: 256 → 128 (pruning rate: 50.0 %)
All non-MoE tensors (attention, layernorms, embeddings, lm_head, MTP heads if any) are bit-identical to the source model.
gate.weight and e_score_correction_bias per MoE layer are row-sliced to the kept experts; per-expert tensors of dropped experts are absent; kept experts are renumbered contiguously to 0..127.
top_k = num_experts_per_tok is unchanged (8).

Method (one-paragraph)

We run a small calibration set (64 prompts spanning Nokia AI4Code, general English Q&A, multilingual, and reasoning) through the unpruned source model and hook every MoE layer's router. Per layer, we accumulate each expert's selected probability mass — the post-sigmoid routing weight that the expert receives, summed over all calibration tokens that selected it in their top-8. We keep the top-K by this score per layer (uniform K) and atomically slice the on-disk per-expert tensors. No gradients, no fine-tuning.

Layer-level statistics

Layers covered: 62
Tokens per layer (calibration): 1,851
Calibration prompts by bucket: {"ai4code": 1008, "general_en": 416, "reasoning": 257, "multilingual": 170}
Median per-layer "kept-min vs drop-max" routing-mass gap: +0.0281 (positive = clean separation between the kept and dropped experts; close to zero or negative = experts of similar utility, expect more quality risk)

Intended use

Production-style serving of the source model's domain (Nokia / Merlin AI4Code plus general English) at reduced HBM footprint. Expect graceful quality degradation versus the unpruned source on tasks well-covered by the calibration mix; quality on out-of-distribution domains may drop further.

Limitations

Training-free: no fine-tune recovery, no distillation, no merge.
Uniform K per layer: late layers may tolerate more pruning than early ones, unexploited here.
Calibration mix is small (64 prompts). Domain coverage is biased toward the included buckets.

Files

config.json, model-NNNNN-of-NNNNN.safetensors (FP8), model.safetensors.index.json, tokenizer, custom modeling_minimax_m2.py + configuration_minimax_m2.py, and expert_prune_plan.json (full record of which experts were kept per layer).

Loading

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("morriszjm/MiniMax-M2.5-128e", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "morriszjm/MiniMax-M2.5-128e", trust_remote_code=True,
    torch_dtype=torch.bfloat16, device_map="auto",
)

For vLLM serving, pass --trust-remote-code and (on multi-GPU) match --data-parallel-size to the EP topology you compiled the K against.