NLP Lineage

11 January 2024 · Language modelling

DeepSeekMoE

DeepSeekMoE: Towards Ultimate Expert Specialization

D. Dai, C. Deng, C. Zhao, et al. (17 authors)

Date
11 January 2024
What the date marks
arXiv submission
Kind
Architectural idea
Organisation
DeepSeek
Access
MIT code, DeepSeek model licence
Threads
In the guided tour
Chapter 8. The weights get out
Learning phases
8. Training from scratch

A milestone: 3 later milestones descend directly from it, and its descendants reach 7 of the 10 threads. How levels are decided

Read the primary source (arxiv.org)

Full summary in progress. Until then, this page shows the milestone's record from the research data.

What it changed

Before it

MoE models used a handful of large experts with top-k routing.

The problem

Coarse experts duplicate common knowledge and specialise poorly.

What it introduced

Many finer-grained experts plus a few always-on shared experts to hold common knowledge.

Why it matters

The architecture behind DeepSeek V2, V3 and R1, and the most copied MoE design of the current era.

What modern models inherit

Fine-grained plus shared experts as the modern MoE layout.

What came next

DeepSeek-V2's latent attention, then V3 and R1.

Before and after

The milestones it descends from, and the milestones that descend from it. Work developed alongside it is not counted. Weak links are left out unless you include them.

Before: where it came from

DeepSeekMoE has 30 documented ancestors across 5 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 11 turning points among them.

DeepSeekMoE has 30 documented ancestors across 5 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

DeepSeekMoE has 31 recorded ancestors, including indirect ones, across 5 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 11 turning points among them.

DeepSeekMoE has 31 recorded ancestors, including indirect ones, across 5 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2024DeepSeekMoEThis milestone.
  2. 2021Switch TransformerDeepSeekMoE was influenced by thisChapter 5. Scale becomes a science
  3. 2020GShard Turning pointDeepSeekMoE built directly on thisChapter 5. Scale becomes a science
  4. 2019T5 Turning pointSwitch Transformer built directly on thisChapter 4. Pretrain once, use everywhere
  5. 2019C4 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  6. 2019GPT-2 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  7. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  8. 2018BERT Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  9. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  10. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  11. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  12. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  13. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  14. 2017Transformer Turning pointT5 built directly on thisChapter 3. Attention, then only attention
  15. 2017Sparse MoE Turning pointDeepSeekMoE was influenced by thisChapter 3. Attention, then only attention
  16. 2016Layer normTransformer built directly on this
    GPT-2 built directly on this    Chapter 3. Attention, then only attention
  17. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  18. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  19. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  20. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  21. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  22. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  23. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  24. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  25. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  26. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  27. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  28. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  29. 1997LSTM Turning pointSparse MoE built directly on thisChapter 1. Language as probability
  30. 1991Mixture of expertsSparse MoE built directly on thisChapter 1. Language as probability
  31. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2024DeepSeekMoEThis milestone.
  2. 2021Switch TransformerDeepSeekMoE was influenced by thisChapter 5. Scale becomes a science
  3. 2020GShard Turning pointDeepSeekMoE built directly on thisChapter 5. Scale becomes a science
  4. 2019T5 Turning pointSwitch Transformer built directly on thisChapter 4. Pretrain once, use everywhere
  5. 2019C4 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  6. 2019GPT-2 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  7. 2019WebTextGPT-2 was influenced by this
    C4 was influenced by this (indirect influence)    Chapter 4. Pretrain once, use everywhere
  8. 2018BERT Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  9. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  10. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  11. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  12. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  13. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  14. 2017Transformer Turning pointT5 built directly on thisChapter 3. Attention, then only attention
  15. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  16. 2017Sparse MoE Turning pointDeepSeekMoE was influenced by thisChapter 3. Attention, then only attention
  17. 2016Layer normTransformer built directly on this
    GPT-2 built directly on this    Chapter 3. Attention, then only attention
  18. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  19. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  20. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  21. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  22. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  23. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  24. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  25. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  26. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  27. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  28. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  29. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  30. 1997LSTM Turning pointSparse MoE built directly on thisChapter 1. Language as probability
  31. 1991Mixture of expertsSparse MoE built directly on thisChapter 1. Language as probability
  32. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

DeepSeekMoE has 10 documented descendants across 4 chapters, reaching forward to Qwen3.8 in 2026. Showing its direct descendants.

DeepSeekMoE has 10 documented descendants across 4 chapters, reaching forward to Qwen3.8 in 2026. Showing all of them.

DeepSeekMoE has 13 recorded descendants, including indirect ones, across 4 chapters, reaching forward to Qwen3.8 in 2026. Showing its direct descendants.

DeepSeekMoE has 13 recorded descendants, including indirect ones, across 4 chapters, reaching forward to Qwen3.8 in 2026. Showing all of them.

  1. 2024DeepSeekMoEThis milestone.
  2. 2024DeepSeek-V2 / MLAbuilt directly on DeepSeekMoEChapter 9. Making it run anywhere
  3. 2024DeepSeek-V3built directly on DeepSeekMoEChapter 8. The weights get out
  4. 2025DeepSeek-R1built directly on DeepSeek-V3Chapter 12. Thinking longer
  5. 2025Qwen3was influenced by DeepSeek-R1Chapter 8. The weights get out
  6. 2025Kimi K2was influenced by DeepSeekMoEChapter 8. The weights get out
  7. 2025DeepSeek-V3.2was influenced by DeepSeek-V2 / MLA
    built directly on DeepSeek-V3    Chapter 10. Reading more than you remember
  8. 2026Qwen3.5built directly on Qwen3Chapter 8. The weights get out
  9. 2026DeepSeek-V4was influenced by DeepSeek-V3Chapter 8. The weights get out
  10. 2026Kimi K3built directly on Kimi K2Chapter 8. The weights get out
  11. 2026Qwen3.8built directly on Qwen3.5Chapter 8. The weights get out
  1. 2024DeepSeekMoEThis milestone.
  2. 2024DeepSeek-V2 / MLAbuilt directly on DeepSeekMoEChapter 9. Making it run anywhere
  3. 2024DeepSeek-V3built directly on DeepSeekMoEChapter 8. The weights get out
  4. 2025DeepSeek-R1built directly on DeepSeek-V3Chapter 12. Thinking longer
  5. 2025Qwen3was influenced by DeepSeek-R1Chapter 8. The weights get out
  6. 2025Kimi K2was influenced by DeepSeekMoEChapter 8. The weights get out
  7. 2025gpt-osswas influenced by DeepSeek-R1 (indirect influence)Chapter 8. The weights get out
  8. 2025GLM-4.5was influenced by DeepSeek-R1 (indirect influence)Chapter 8. The weights get out
  9. 2025DeepSeek-V3.2was influenced by DeepSeek-V2 / MLA
    built directly on DeepSeek-V3    Chapter 10. Reading more than you remember
  10. 2026Qwen3.5built directly on Qwen3Chapter 8. The weights get out
  11. 2026DeepSeek-V4was influenced by DeepSeek-V3Chapter 8. The weights get out
  12. 2026GLM-5.2built directly on GLM-4.5Chapter 8. The weights get out
  13. 2026Kimi K3built directly on Kimi K2Chapter 8. The weights get out
  14. 2026Qwen3.8built directly on Qwen3.5Chapter 8. The weights get out

Direct links and their sources

Where it came from

Built directly on

Influenced by

What it led to

Built on directly by

Influenced

Terms