NLP Lineage

6 May 2024 · Efficiency and infrastructure

DeepSeek-V2 / MLA

DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model

DeepSeek-AI

Date
6 May 2024
What the date marks
arXiv submission
Kind
Model
Organisation
DeepSeek
Access
MIT code, DeepSeek model licence
Threads
In the guided tour
Chapter 9. Making it run anywhere
Learning phases
1. Transformer architecture; 8. Training from scratch

A milestone: 2 later milestones descend directly from it, and its descendants reach 5 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

Grouped-query attention was the standard way to shrink the KV cache.

The problem

Even grouped keys and values dominate memory during long-context inference.

What it introduced

Multi-head latent attention: compress keys and values into a low-rank latent vector and cache that instead.

Why it matters

A far larger cache reduction than GQA, and the reason DeepSeek could serve long context so cheaply.

What modern models inherit

Latent KV compression as the current frontier of inference efficiency.

What came next

DeepSeek-V3 and R1 use it; later sparse attention built on top.

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

DeepSeek-V2 / MLA has 35 documented ancestors across 8 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 13 turning points among them.

DeepSeek-V2 / MLA has 35 documented ancestors across 8 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

DeepSeek-V2 / MLA has 36 recorded ancestors, including indirect ones, across 8 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 13 turning points among them.

DeepSeek-V2 / MLA has 36 recorded ancestors, including indirect ones, across 8 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2024DeepSeek-V2 / MLAThis milestone.
  2. 2024DeepSeekMoEDeepSeek-V2 / MLA built directly on thisChapter 8. The weights get out
  3. 2023Grouped-query attention Turning pointDeepSeek-V2 / MLA was influenced by thisChapter 9. Making it run anywhere
  4. 2021RoPE Turning pointDeepSeek-V2 / MLA built directly on thisChapter 10. Reading more than you remember
  5. 2021Switch TransformerDeepSeekMoE was influenced by thisChapter 5. Scale becomes a science
  6. 2020GShard Turning pointDeepSeekMoE built directly on thisChapter 5. Scale becomes a science
  7. 2019Multi-query attentionDeepSeek-V2 / MLA was influenced by thisChapter 9. Making it run anywhere
  8. 2019T5 Turning pointSwitch Transformer built directly on thisChapter 4. Pretrain once, use everywhere
  9. 2019C4 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  10. 2019GPT-2 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  11. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  12. 2019Transformer-XLRoPE was influenced by thisChapter 4. Pretrain once, use everywhere
  13. 2018BERT Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  14. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  15. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  16. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  17. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  18. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  19. 2017Transformer Turning pointMulti-query attention built directly on this
    RoPE built directly on this    Chapter 3. Attention, then only attention
  20. 2017Sparse MoE Turning pointDeepSeekMoE was influenced by thisChapter 3. Attention, then only attention
  21. 2016Layer normTransformer built directly on thisChapter 3. Attention, then only attention
  22. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  23. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  24. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  25. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  26. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  27. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  28. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  29. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  30. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  31. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  32. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  33. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  34. 1997LSTM Turning pointSparse MoE built directly on thisChapter 1. Language as probability
  35. 1991Mixture of expertsSparse MoE built directly on thisChapter 1. Language as probability
  36. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2024DeepSeek-V2 / MLAThis milestone.
  2. 2024DeepSeekMoEDeepSeek-V2 / MLA built directly on thisChapter 8. The weights get out
  3. 2023Grouped-query attention Turning pointDeepSeek-V2 / MLA was influenced by thisChapter 9. Making it run anywhere
  4. 2021RoPE Turning pointDeepSeek-V2 / MLA built directly on thisChapter 10. Reading more than you remember
  5. 2021Switch TransformerDeepSeekMoE was influenced by thisChapter 5. Scale becomes a science
  6. 2020GShard Turning pointDeepSeekMoE built directly on thisChapter 5. Scale becomes a science
  7. 2019Multi-query attentionDeepSeek-V2 / MLA was influenced by thisChapter 9. Making it run anywhere
  8. 2019T5 Turning pointSwitch Transformer built directly on thisChapter 4. Pretrain once, use everywhere
  9. 2019C4 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  10. 2019GPT-2 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  11. 2019WebTextGPT-2 was influenced by this
    C4 was influenced by this (indirect influence)    Chapter 4. Pretrain once, use everywhere
  12. 2019Transformer-XLRoPE was influenced by thisChapter 4. Pretrain once, use everywhere
  13. 2018BERT Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  14. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  15. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  16. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  17. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  18. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  19. 2017Transformer Turning pointMulti-query attention built directly on this
    RoPE built directly on this    Chapter 3. Attention, then only attention
  20. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  21. 2017Sparse MoE Turning pointDeepSeekMoE was influenced by thisChapter 3. Attention, then only attention
  22. 2016Layer normTransformer built directly on thisChapter 3. Attention, then only attention
  23. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  24. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  25. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  26. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  27. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  28. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  29. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  30. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  31. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  32. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  33. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  34. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  35. 1997LSTM Turning pointSparse MoE built directly on thisChapter 1. Language as probability
  36. 1991Mixture of expertsSparse MoE built directly on thisChapter 1. Language as probability
  37. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

DeepSeek-V2 / MLA has 7 documented descendants across 3 chapters, reaching forward to Qwen3.8 in 2026. Showing its direct descendants.

DeepSeek-V2 / MLA has 7 documented descendants across 3 chapters, reaching forward to Qwen3.8 in 2026. Showing all of them.

DeepSeek-V2 / MLA has 12 recorded descendants, including indirect ones, across 3 chapters, reaching forward to Qwen3.8 in 2026. Showing its direct descendants.

DeepSeek-V2 / MLA has 12 recorded descendants, including indirect ones, across 3 chapters, reaching forward to Qwen3.8 in 2026. Showing all of them.

  1. 2024DeepSeek-V2 / MLAThis milestone.
  2. 2024DeepSeek-V3built directly on DeepSeek-V2 / MLAChapter 8. The weights get out
  3. 2025DeepSeek-R1built directly on DeepSeek-V3Chapter 12. Thinking longer
  4. 2025Qwen3was influenced by DeepSeek-R1Chapter 8. The weights get out
  5. 2025DeepSeek-V3.2was influenced by DeepSeek-V2 / MLAChapter 10. Reading more than you remember
  6. 2026Qwen3.5built directly on Qwen3Chapter 8. The weights get out
  7. 2026DeepSeek-V4built directly on DeepSeek-V3.2
    was influenced by DeepSeek-V3    Chapter 8. The weights get out
  8. 2026Qwen3.8built directly on Qwen3.5Chapter 8. The weights get out
  1. 2024DeepSeek-V2 / MLAThis milestone.
  2. 2024DeepSeek-V3built directly on DeepSeek-V2 / MLAChapter 8. The weights get out
  3. 2025DeepSeek-R1built directly on DeepSeek-V3Chapter 12. Thinking longer
  4. 2025Qwen3was influenced by DeepSeek-R1Chapter 8. The weights get out
  5. 2025Kimi K2was influenced by DeepSeek-V3 (indirect influence)Chapter 8. The weights get out
  6. 2025gpt-osswas influenced by DeepSeek-R1 (indirect influence)Chapter 8. The weights get out
  7. 2025GLM-4.5was influenced by DeepSeek-R1 (indirect influence)Chapter 8. The weights get out
  8. 2025DeepSeek-V3.2was influenced by DeepSeek-V2 / MLAChapter 10. Reading more than you remember
  9. 2026Qwen3.5built directly on Qwen3Chapter 8. The weights get out
  10. 2026DeepSeek-V4built directly on DeepSeek-V3.2
    was influenced by DeepSeek-V3    Chapter 8. The weights get out
  11. 2026GLM-5.2built directly on GLM-4.5Chapter 8. The weights get out
  12. 2026Kimi K3built directly on Kimi K2Chapter 8. The weights get out
  13. 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