AIFEATURE

What Is MoE (Mixture of Experts)? Why It Makes Large Language Models Cheaper and Faster

N
NathanTechnology Editor · Technical Lead
Published · Updated
Mixture of Experts (MoE) architecture separates a large language model's total parameter count from its per-token activated parameter count, letting models such as Qwen3 235B-A22B run inference at roughly the speed of a much smaller dense model while keeping a large model's knowledge capacity — the trade-off is that memory requirements still scale with total parameters, not activated ones.

What Is a Mixture of Experts (MoE) Architecture?

Mixture of Experts (MoE) is a large language model architecture in which a model is built from many "expert" sub-networks, but only a small subset of those experts is activated on each inference pass rather than the full parameter setCITE:E1.

The defining feature of MoE is the separation between a model's "total parameters" and its "activated parameters": Qwen3 235B-A22B carries 235 billion total parameters but activates only 22 billion per pass, while DeepSeek V4-Pro carries 1.6 trillion total parameters but activates only 49 billion per passCITE:E2. This split allows overall model capacity to grow far beyond what the per-token compute cost alone would suggest.

Why Does MoE Make Large Model Inference Both Cheaper and Faster?

Inference speed in an MoE model tracks its activated-parameter count, not its total-parameter count, so a 235-billion-parameter MoE model can run at roughly the speed of a 22-billion-parameter dense model while retaining a large model's knowledge capacityCITE:E3.

What Advantage Does MoE Offer for Local Deployment?

MoE benefits local deployment because system memory must hold the full set of parameters (weights), while compute bandwidth only needs to serve the activated subset, so devices with large memory pools — such as Apple Silicon Macs with unified memory — can hold large MoE models in memory and still run them quicklyCITE:E4.

What Are MoE's Limitations and Trade-offs?

MoE reduces compute cost per token but does not reduce memory footprint: fitting a model's full total-parameter count still requires memory sized to that total, and trillion-parameter-class models such as Kimi K2 and DeepSeek V4-Pro exceed single-machine memory limits even after 4-bit quantization, meaning they cannot run locallyCITE:E5. This is the architecture's core trade-off — it saves compute, not memory.

Why Do Mainstream Open-Weight Models in 2026 Adopt MoE?

Most mainstream open-weight models released in 2026 — including DeepSeek V4, Qwen3, Kimi K2, and GLM — use MoE architecture, reflecting an industry-wide approach to reconciling large-model capability with controlled inference costCITE:E6.

Total vs. Activated Parameters, By Model

ModelTotal ParametersActivated Parameters
Qwen3 235B-A22B235B22BCITE:E2
DeepSeek V4-Pro1.6T49BCITE:E2
Kimi K2Trillion-parameter classExceeds single-machine memory even at 4-bit quantizationCITE:E5

What this means: The same activated-parameter mechanism that lets Qwen3 235B-A22B and DeepSeek V4-Pro run inference at the speed of much smaller dense modelsCITE:E2CITE:E3 is also what makes them fit on memory-rich local devices like Apple Silicon MacsCITE:E4. But that mechanism only addresses compute — it does nothing for memory, which is why Kimi K2 and DeepSeek V4-Pro, both trillion-parameter-class, still exceed single-machine memory limits after 4-bit quantizationCITE:E5. That compute/memory split is the specific trade-off behind why DeepSeek V4, Qwen3, Kimi K2, and GLM all converged on MoE as open-weight releases in 2026CITE:E6.

📊 Evidence

FAQ

What Is a Mixture of Experts (MoE) Architecture?

Mixture of Experts (MoE) is a large language model architecture in which a model is built from many "expert" sub-networks, but only a small subset of those expe…

Why Does MoE Make Large Model Inference Both Cheaper and Faster?

Inference speed in an MoE model tracks its activated-parameter count, not its total-parameter count, so a 235-billion-parameter MoE model can run at roughly the…

What Advantage Does MoE Offer for Local Deployment?

MoE benefits local deployment because system memory must hold the full set of parameters (weights), while compute bandwidth only needs to serve the activated su…

What Are MoE's Limitations and Trade-offs?

MoE reduces compute cost per token but does not reduce memory footprint: fitting a model's full total-parameter count still requires memory sized to that total,…

📎 Sources

  1. effectstory.com
  2. effectstory.com

Related data

Author's TakeNathan

The real story here isn't that MoE makes models fast — it's that it decouples two variables that used to move together: total capacity and per-token compute. Qwen3 235B-A22B activates roughly 22 of its 235 billion parameters per pass, and DeepSeek V4-Pro activates roughly 49 of its 1.6 trillion — in both cases a small fraction of the total, which is why the activated-parameter figure, not the headline parameter count, is the number to watch as more MoE models ship. But the trade-off documented here is memory, not compute: Kimi K2 and DeepSeek V4-Pro still exceed single-machine memory even after 4-bit quantization. The metric worth tracking next is whether unified-memory devices like Apple Silicon Macs — which can already hold Qwen3-class MoE models — scale memory capacity fast enough to keep pace with total-parameter growth in future MoE releases.

N
NathanTechnology Editor · Technical Lead

Related

BRIEF

SpaceX and NVIDIA Plan to Launch Orbital AI Data Centers Starting Next Year, Musk Says

SpaceX plans to launch its first NVIDIA-chip-powered AI satellites, Starmind AI1, in the fourth quarter of 2027, reaching large-scale deployment by 2028 — at least a year ahead of the original schedule. The company has also filed with the FCC for a network of up to 1 million satellites, while Taiwanese suppliers Unitech and Sesoda report rising satellite-related revenue and order shares tied to the buildout.

林紀旭 James Lin ·
BRIEF

NVIDIA Launches Jetson Orin Nano 2 Robotics Computer, Targets Entry-Level Edge AI

NVIDIA announced the Jetson Orin Nano 2 robotics computer on August 25, 2026, packing 78 TOPS of AI compute, 8GB of memory, and an 8-core Arm CPU while doubling inference performance and cutting power draw 40% versus its predecessor. Modules and developer kits ship in the first half of 2027, targeting a robotics developer base NVIDIA says already tops 3 million.

Nathan ·