Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

BitsMoE: Efficient Spectral Energy-Guided Bit Allocation for MoE LLM Quantization

About

Mixture-of-Experts (MoE) large language models reduce per-token computation through sparse expert activation, but their deployment remains memory-intensive because all expert weights must be kept resident in memory. Existing MoE compression methods struggle in the ultra-low-bit regime: pruning irreversibly removes model capacity, while coarse-grained quantization fails to allocate bits according to heterogeneous expert and weight-direction importance. We propose BitsMoE, a spectral-energy-guided bit-allocation framework for MoE LLM quantization. BitsMoE decomposes each MoE layer by SVD into a shared basis and expert-specific spectral factors, retaining the shared basis without quantization to preserve common cross-expert structure and using the expert-specific factors as fine-grained quantization units. To determine the bit-width of each unit, BitsMoE formulates spectrum-wise mixed-precision quantization as an activation-aware reconstruction surrogate and solves an integer linear program that minimizes estimated reconstruction loss under a fixed bit budget. Experiments across multiple MoE LLMs show that BitsMoE substantially reduces downstream task accuracy degradation in ultra-low-bit regimes. Under 2-bit quantization on Qwen3-30B-A3B-Base, BitsMoE accelerates quantization by 12.3$\times$, improves average accuracy by 27.83 percentage points, and increases decoding speed by 1.76$\times$ over GPTQ. Our model and code are publicly available at https://github.com/zjiayu064/BitsMoE.

Jiayu Zhao, Zihan Teng, Minhao Fan, Tianrui Ma, Wentao Ren, Song Chen, Weichen Liu• 2026

Related benchmarks

TaskDatasetResultRank
Language ModelingC4
Perplexity9.38
72
Zero-shot General EvaluationZero-shot Task Suite (HellaSwag, MathQA, MMLU, OpenBookQA, WinoGrande, GSM8K, HumanEval)
HellaSwag Accuracy80.93
31
Inference EfficiencyMoE LLMs DSV2-16B, QW3-30B, QW3-80B-I
Decode Speed (tokens/sec)12.46
9
General Language Understanding and Reasoning7-Task Evaluation Suite (HellaSwag, MathQA, MMLU, OpenBookQA, WinoGrande, GSM8K, HumanEval)
Average Accuracy61.91
8
Showing 4 of 4 rows

Other info

Follow for update