NLP Lineage

11 January 2021 · Scale and data

Switch Transformer

Switch Transformers: Scaling to Trillion Parameter Models

W. Fedus, B. Zoph, N. Shazeer

Date
11 January 2021
What the date marks
arXiv submission
Kind
Model
Organisation
Google
Access
Apache 2.0
Other dates
JMLR 2022
Threads
In the guided tour
Chapter 5. Scale becomes a science
Learning phases
8. Training from scratch

A milestone: 3 later milestones descend directly from it, and its descendants reach 8 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

Mixture-of-experts routing sent each token to several experts, with fiddly balancing losses.

The problem

Multi-expert routing was complex, unstable and communication-heavy.

What it introduced

Route each token to exactly one expert. Simpler, faster, and stable to a trillion parameters.

Why it matters

It made sparse scaling simple enough to copy, which is why MoE is now standard at the frontier.

What modern models inherit

Top-1 style routing and the parameter-count-versus-active-parameters distinction.

What came next

Mixtral, DeepSeekMoE, Llama 4 and Kimi all descend from this simplification.

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

Switch Transformer has 29 documented ancestors across 5 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 11 turning points among them.

Switch Transformer has 29 documented ancestors across 5 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

Switch Transformer has 30 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.

Switch Transformer has 30 recorded ancestors, including indirect ones, across 5 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

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

After: what it led to

Switch Transformer has 14 documented descendants across 4 chapters, reaching forward to Qwen3.8 in 2026. Showing its direct descendants.

Switch Transformer has 14 documented descendants across 4 chapters, reaching forward to Qwen3.8 in 2026. Showing all of them.

Switch Transformer has 17 recorded descendants, including indirect ones, across 4 chapters, reaching forward to Qwen3.8 in 2026. Showing its direct descendants.

Switch Transformer has 17 recorded descendants, including indirect ones, across 4 chapters, reaching forward to Qwen3.8 in 2026. Showing all of them.

  1. 2021Switch TransformerThis milestone.
  2. 2023Mixtral 8x7Bbuilt directly on Switch TransformerChapter 8. The weights get out
  3. 2024DeepSeekMoEwas influenced by Switch TransformerChapter 8. The weights get out
  4. 2024DeepSeek-V2 / MLAbuilt directly on DeepSeekMoEChapter 9. Making it run anywhere
  5. 2024DeepSeek-V3built directly on DeepSeekMoEChapter 8. The weights get out
  6. 2025DeepSeek-R1built directly on DeepSeek-V3Chapter 12. Thinking longer
  7. 2025Llama 4was influenced by Switch TransformerChapter 8. The weights get out
  8. 2025Qwen3was influenced by DeepSeek-R1Chapter 8. The weights get out
  9. 2025Kimi K2was influenced by DeepSeekMoEChapter 8. The weights get out
  10. 2025DeepSeek-V3.2was influenced by DeepSeek-V2 / MLA
    built directly on DeepSeek-V3    Chapter 10. Reading more than you remember
  11. 2026Qwen3.5built directly on Qwen3Chapter 8. The weights get out
  12. 2026Muse Sparksucceeded Llama 4Chapter 8. The weights get out
  13. 2026DeepSeek-V4was influenced by DeepSeek-V3Chapter 8. The weights get out
  14. 2026Kimi K3built directly on Kimi K2Chapter 8. The weights get out
  15. 2026Qwen3.8built directly on Qwen3.5Chapter 8. The weights get out
  1. 2021Switch TransformerThis milestone.
  2. 2023Mixtral 8x7Bbuilt directly on Switch TransformerChapter 8. The weights get out
  3. 2024DeepSeekMoEwas influenced by Switch TransformerChapter 8. The weights get out
  4. 2024DeepSeek-V2 / MLAbuilt directly on DeepSeekMoEChapter 9. Making it run anywhere
  5. 2024DeepSeek-V3built directly on DeepSeekMoEChapter 8. The weights get out
  6. 2025DeepSeek-R1built directly on DeepSeek-V3Chapter 12. Thinking longer
  7. 2025Llama 4was influenced by Switch TransformerChapter 8. The weights get out
  8. 2025Qwen3was influenced by DeepSeek-R1Chapter 8. The weights get out
  9. 2025Kimi K2was influenced by DeepSeekMoEChapter 8. The weights get out
  10. 2025gpt-osswas influenced by DeepSeek-R1 (indirect influence)Chapter 8. The weights get out
  11. 2025GLM-4.5was influenced by DeepSeek-R1 (indirect influence)Chapter 8. The weights get out
  12. 2025DeepSeek-V3.2was influenced by DeepSeek-V2 / MLA
    built directly on DeepSeek-V3    Chapter 10. Reading more than you remember
  13. 2026Qwen3.5built directly on Qwen3Chapter 8. The weights get out
  14. 2026Muse Sparksucceeded Llama 4Chapter 8. The weights get out
  15. 2026DeepSeek-V4was influenced by DeepSeek-V3Chapter 8. The weights get out
  16. 2026GLM-5.2built directly on GLM-4.5Chapter 8. The weights get out
  17. 2026Kimi K3built directly on Kimi K2Chapter 8. The weights get out
  18. 2026Qwen3.8built directly on Qwen3.5Chapter 8. The weights get out

Direct links and their sources

Where it came from

Built directly on

What it led to

Built on directly by

Influenced

Terms