How LatentMoE Reshapes the MoE Paradigm
青稞AI · wechat · 2026-07-19
Using Kimi K3 and NVIDIA Nemotron3 Super as examples, this article explains why LatentMoE could be the key paradigm for the next generation of MoE models.
The core argument is that traditional MoE bottlenecks aren't just about the number of experts. When routing tokens to experts, large-dimension weights must be read and all-to-all communication is required—costs that scale linearly as the model grows. LatentMoE addresses this by first projecting tokens from their original dimension d into a smaller latent space ℓ, performing routing and expert computations in this lower-dimensional space, and then projecting back. This allows it to:
- Reduce expert weight reads from d×m to ℓ×m
- Proportionally decrease all-to-all communication volume
- Activate more experts and expand expert combination space under the same inference cost
The article details two case studies:
- Nemotron3 Super: Features 512 experts, Top-22 routing, ℓ = 1024, and combines Mamba-2, minimal attention, and MTP for inference acceleration.
- Kimi K3: Features 896 experts, Top-16 routing, and an extremely high sparsity ratio. It introduces QuantileBalancing—using quantiles instead of heuristic hyperparameters for load balancing—to maintain training stability at a 2.8T scale.
The author concludes that while NVIDIA leans towards "efficiency first" and Kimi pushes for "extreme scale," both are essentially compressing expert computation dimensions to trade saved resources for greater expert combinatorial capacity. The article also looks ahead to adaptive compression ratios, dynamic Top-K routing, and incorporating structured sparsity into architectural design.
More from Infra
- SF Compute founder: buying compute is 'an absolutely awful experience' right now — IgorCarron · 2026-09-11
- SmolVM open-sources persistent computer infrastructure for agents that outlive chat sessions — aniketmaurya · 2026-09-11
- PyTorch Day Korea 2026 launches first offline conf, CFP closes Sept 13 — PyTorch · 2026-09-11
- Local LLM server dilemma: 4x CMP-170HX (price up 53% in 20 days) vs Mac Studio M5 Ultra — rumboll · 2026-09-11
- llama.cpp lands Flash Attention tuning for RDNA4, big prefill gains on AMD — pmttyji · 2026-09-11
- Your p99 latency benchmark may be lying: a deep dive into coordinated omission — Franc0Fernand0 · 2026-09-11