Back to the episode map

Research Note

Mixture of Experts Architecture Record

A mixture-of-experts transformer contains multiple expert sub-networks and a router that selects a subset for each token or representation. Sparse activation can provide

Aug 4, 20262 min readBy Dalton Anderson

Mixture of Experts Architecture Record

Definition

A mixture-of-experts transformer contains multiple expert sub-networks and a router that selects a subset for each token or representation. Sparse activation can provide more total model capacity without activating every parameter for every token.

DeepSeek-V3 is the episode's concrete example:

https://github.com/deepseek-ai/DeepSeek-V3

https://arxiv.org/abs/2412.19437

The authors describe 671 billion total parameters and 37 billion activated per token. These figures belong to the specified architecture and do not mean that the full model stores only 37 billion parameters.

Parameter terms

Total parameters describe the full learned parameter set. Active parameters describe the subset participating in computation for a token under the routing design.

Neither number alone states memory footprint, bandwidth, latency, throughput, energy, quality, or cost.

Weights for inactive experts may still need to reside in memory across the serving system. Routing can distribute tokens unevenly. Expert placement can require communication across devices and nodes.

System map

LayerQuestion
RouterHow are experts selected, and what happens when routing is wrong or unstable?
CapacityHow many experts are available and active?
BalanceHow are hot experts and underused experts handled?
MemoryWhere do all weights, optimizer states, caches, and activations live?
CommunicationHow do tokens move across devices or nodes to their experts?
ParallelismHow are data, tensor, pipeline, and expert parallelism combined?
TrainingWhat losses, precision, scheduling, and failure controls are used?
ServingWhat batch, sequence, quantization, caching, and routing conditions apply?
OperationsWhat monitoring, failure, scaling, and capacity controls are required?

Tradeoff boundary

Sparse activation can reduce arithmetic relative to a dense model with the same total parameter count. It does not make routing, memory, network, synchronization, load balancing, serving, or operations free.

Delivered performance depends on hardware and software co-design, precision, batch, sequence length, network topology, expert placement, kernels, compiler, runtime, and workload.

Architecture alone cannot explain every reported efficiency or quality result. Data, objectives, training process, implementation, hardware, and evaluation also matter.

Publication language

Avoid "only 37 billion parameters." Say that the cited architecture has 671 billion total parameters and activates 37 billion per token under the authors' design.

Avoid presenting active-parameter count as direct proof of a dollar cost or hardware requirement.

Sources

Follow the evidence.

  1. github.com: LICENSEgithub.com
  2. bis.gov: commerce strengthens restrictions advanced computing semiconductors enhance foundry due diligence preventbis.gov
  3. arxiv.org: 2501arxiv.org
  4. NIST AI Risk Management Frameworknist.gov
  5. daltonanderson.ghost.io: deepseek vs nvidia the future of ai chip economicsdaltonanderson.ghost.io
  6. investor.nvidia.com: defaultinvestor.nvidia.com
  7. api-docs.deepseek.comapi-docs.deepseek.com
  8. bis.gov: 740bis.gov
  9. daltonanderson.net: deepseek vs nvidia the future of ai chip economicsdaltonanderson.net
  10. github.com: DeepSeek R1github.com
  11. open.spotify.com: 6jLI1bNwyoxI449vXJXzBVopen.spotify.com
  12. youtu.be: Qp24TkfT9XEyoutu.be
  13. bis.gov: 742bis.gov
  14. bis.gov: department commerce revises license review policy semiconductors exported chinabis.gov
  15. arxiv.org: 2412arxiv.org
  16. docs.nvidia.com: cudadocs.nvidia.com
  17. github.com: DeepSeek V3github.com
Mixture of Experts Architecture Record