Revision History
Mixture of Experts (MoE) · 20 revisions
Sizes are character counts of the article source. The signed number is the change from the previous revision.
Recent edit summaries
Detailed summaries recorded by editors. Generic maintenance summaries are omitted here; every recorded revision remains below. Dates describe the edit, not necessarily the event it covers.
- Expanded training comparisons, quantile load balancing, execution tradeoffs, and systems-aware scaling
Version 20 · Nov 8, 2026, 08:29 AM
- Update: added a video.
Version 19 · Nov 1, 2026, 08:05 AM
- Update: expanded historical development (noisy top-k gating, GShard, Switch, GLaM, ST-MoE, Mixtral, DeepSeekMoE) with a milestones table; added gate formulas, auxiliary-loss-free balancing, sparse upcycling results, fine-grained and shared experts, MoE scaling laws, 2023-2026 expert-count trends, and vendor statements on Gemini 1.5 and 2.5; audited the open-weight model table against model cards and config files.
Version 18 · Oct 31, 2026, 03:48 AM
- Update: inference subsections (memory-bound decoding, serving-scale expert parallelism, fused kernels, low-precision experts, offloading, speculative decoding) and a table of open-weight MoE models
Version 17 · Oct 29, 2026, 09:33 PM
- Added an explanatory diagram with caption and image credits
Version 16 · Oct 4, 2026, 06:01 PM