mlsanityData to 7 Oct 2026

The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction

arXiv 2609.18063 ↗First posted 16 Sep 2026cs.AIcited by 0

On the map: Mixture-of-experts → Efficient MoE training and inference systems

Abstract (arXiv)

Mixture-of-experts (MoE) inference on consumer hardware is bounded by weight memory: a 35B-class model is 19.5GB at 4-bit, and sparsity shrinks the compute per token, not the bytes that must be held. Naive offloading to SSD does not help on its own, because layer N+1's experts must be chosen before layer N's output exists, so the reads cannot start early enough to hide behind compute. We present Edge0, a streaming MoE inference engine that closes the gap with a prerouter: a per-layer head predicts the next layer's routing one token ahead, and the prediction is consumed as the routing itself, so the staged expert set equals the routed set and nothing is dropped. An unmerged recovery LoRA, trained on the student path, pays back the quality lost to int4 quantization and routing replacement. On a single 24GB machine, Edge0 serves a 35B MoE at 20tok/s inside 3GiB of peak active memory, within a few points of its fp16 teacher on average across five public benchmarks. An 8B tier runs on the same framework, and the framework, checkpoints, and adapters are open source.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Efficient MoE training and inference systems, one of 12 directions in Mixture-of-experts. Builds on 23 of its 32 references that are in the map. First posted on arXiv 16 Sep 2026.

Builds on

23 of 32 references.

Closest ancestors · 3

Same direction, newest first

  1. 2023Fast Inference of Mixture-of-Experts Language Models with Offloading28 Dec
  2. 2023Pre-gated MoE: An Algorithm-System Co-Design for Fast and Scalable Mixture-of-Expert Inference23 Aug
  3. 2022DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale14 Jan

Other references · 20

  1. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,767
  2. 2023Mamba: Linear-Time Sequence Modeling with Selective State Spacescited by 4,616
  3. 2017Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layercited by 3,413
  4. 2023QLoRA: Efficient Finetuning of Quantized LLMscited by 3,355
  5. 2021Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsitycited by 3,014
  6. 2022GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformerscited by 1,838
  7. 2024Mixtral of Expertscited by 1,785
  8. 2020GShard: Scaling Giant Models with Conditional Computation and Automatic Shardingcited by 1,601
  9. 2024DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Modelcited by 1,093
  10. 2023On-Policy Distillation of Language Models: Learning from Self-Generated Mistakescited by 840
  11. 2023FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPUcited by 567
  12. 2023PowerInfer: Fast Large Language Model Serving with a Consumer-grade GPUcited by 210
8 more
  1. 2023MiniLLM: On-Policy Distillation of Large Language Modelscited by 129
  2. 2023MultiLoRA: Democratizing LoRA for Better Multi-Task Learningcited by 32
  3. 2025Every FLOP Counts: Scaling a 300B Mixture-of-Experts LING LLM without Premium GPUscited by 22
  4. 2023PrivateLoRA For Efficient Privacy Preserving LLMcited by 14
  5. 2026Don't Ignore the Tail: Decoupling top-K Probabilities for Efficient Language Model Distillationcited by 7
  6. 2026Post-Trained MoE Can Skip Half Experts via Self-Distillationcited by 4
  7. 2026Tail-Aware Top-k On-Policy Distillationcited by 3
  8. 2026Tiny-Engram: Trigger-Indexed Concept Tables for Generative Visioncited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026FlashMoE: Reducing SSD I/O Bottlenecks via ML-Based Cache Replacement for Mixture-of-Experts Inference on Edge Devices
  2. 2026MoE-SpAc: Efficient MoE Inference Based on Speculative Activation Utility in Heterogeneous Edge Scenarios
  3. 2026SSD-LLaMA: SSD-Native Inference for Trillion-Parameter MoE at 1+ Token/s on a Consumer PC
  4. 2026Mira: Memory-Efficient MoE Inference Using Adaptive Caching and Predictive Expert Staging
  5. 2025Accelerating Mixture-of-Expert Inference with Adaptive Expert Split Mechanism
  6. 2026Does Mixture-of-Experts Actually Help Inference on Consumer and Edge Hardware? An Empirical Study
  7. 2026Efficient Mixture-of-Experts with Speculative Decoding via Expert Coactivation
  8. 2025SSD Offloading for LLM Mixture-of-Experts Weights Considered Harmful in Energy Efficiency