Mixture of experts
A model architecture where multiple specialized expert networks share the work — a routing layer activates only a few experts per input, cutting inference cost while keeping total parameter count high.
Mixture-of-experts (MoE) is the architectural trick that makes frontier-quality models economically feasible. Instead of activating all 100B+ parameters for every token, an MoE model has dozens of specialized expert sub-networks and routes each token to only 2–8 of them. Total parameters stay massive; active parameters per token stay small.
The 2026 frontier is MoE-dominant. Mixtral, DeepSeek-V3, Llama 4, and most leading proprietary models use MoE architectures. The signature pattern: 200B–700B total parameters, 20B–40B active per token, and inference speed comparable to a dense 30B model.
For agent builders, MoE matters because it widens the cost-quality frontier. You can get near-frontier quality at mid-tier inference cost, which makes high-quality agents viable at scale. The trade-off: harder to fine-tune (routing decisions are sensitive), and longer time-to-first-token on cold cache.
Frequently asked
Are all frontier models MoE in 2026?+
Most are. Dense architectures still ship (especially for fine-tuning targets), but the leading frontier models from OpenAI, Anthropic, Google, DeepSeek, and Meta are MoE or hybrid-MoE.
Does MoE affect agent behavior?+
Indirectly. MoE models sometimes show inconsistency on similar inputs because they route to different experts. For agent reliability, this matters less than for chat — agent loops self-correct most routing-induced variation.