Mixture-of-experts routing
MoEQuant
Superseded — cited as a baseline and beaten by newer methods
2 papers critique it · 4 beat it on benchmarks
What papers say
Verbatim critique sentences, each from a paper that cites MoEQuant as a baseline.
While MoEQuant~hu2025moequant attempts to alleviate this issue through expert-balanced self-sampling, such generative calibration approaches may compromise fair comparisons with other baselines.
“MoEQuant moequant uses routing statistics to balance the contributions of each expert during calibration, but its performance remains unsatisfactory under quantization of $ 4$ bits.”
Beaten on benchmarks
Head-to-head results where a newer method reports beating MoEQuant. Values are copied from the source paper's tables — verify against the cited paper.
KBVQ-MoE beats MoEQuant
63.10 vs 25.59
Avg Acc · [DeepseekV2-Lite, 2-bit]
KBVQ-MoE: KLT-guided SVD with Bias-Corrected Vector Quantization for MoE Large Language ModelsBitsMoE beats MoEQuant
61.91 vs 52.21
Avg. Accuracy · [Qwen3-30B-A3B-Base, 2-bit]
BitsMoE: Efficient Spectral Energy-Guided Bit Allocation for MoE LLM QuantizationEAQuant beats MoEQuant
7.47 vs 8.24
C4 Perplexity · [Mixtral-8x7B W3A16]
EAQuant: Enhancing Post-Training Quantization for MoE Models via Expert-Aware OptimizationGEMQ beats MoEQuant
59.49 vs 57.24
0-shot accuracy · [3-16 bits (weight-activation)]
GEMQ: Global Expert-Level Mixed-Precision Quantization for MoE LLMs
What to use instead
Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.
- May 22, 2026
- May 21, 2026
- KBVQ-MoEKBVQ-MoE: KLT-guided SVD with Bias-Corrected Vector Quantization for MoE Large Language ModelsJan 30, 2026
- Oct 13, 2025