NLP Lineage

5 April 2025 · Efficiency and infrastructure

Llama 4

Llama 4

Meta AI

Date
5 April 2025
What the date marks
Model release
Kind
Model
Organisation
Meta AI
Access
Llama 4 Community License
Other dates
Scout and Maverick
Threads
In the guided tour
Chapter 8. The weights get out
Learning phases
6. Local models and deployment

A milestone: 1 later milestone descends directly from it, and its descendants reach 3 of the 10 threads. How levels are decided

Figures quoted here are reported by the organisation that built it, not independently verified.

Read the primary source (ai.meta.com)

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

What it changed

Before it

Llama models were dense and text-first.

The problem

Dense models had become uncompetitive on cost against the open MoE models coming out of China.

What it introduced

Meta's first mixture-of-experts models, natively multimodal, with very long advertised context.

Why it matters

It brought Meta into the MoE era, and turned out to be the last open Llama generation.

What modern models inherit

MoE as the default frontier architecture, open or closed.

What came next

Meta pivoted to the closed Muse Spark line in April 2026.

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

Llama 4 has 50 documented ancestors across 9 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 21 turning points among them.

Llama 4 has 50 documented ancestors across 9 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

Llama 4 has 56 recorded ancestors, including indirect ones, across 9 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 21 turning points among them.

Llama 4 has 56 recorded ancestors, including indirect ones, across 9 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2025Llama 4This milestone.
  2. 2024Llama 3.1 405BLlama 4 succeeded thisChapter 8. The weights get out
  3. 2024Llama 3Llama 3.1 405B built directly on thisChapter 8. The weights get out
  4. 2023Llama 2 Turning pointLlama 3 built directly on thisChapter 8. The weights get out
  5. 2023Position interpolationLlama 3.1 405B was influenced by thisChapter 10. Reading more than you remember
  6. 2023DPOLlama 3.1 405B built directly on thisChapter 6. Teaching models what people want
  7. 2023Grouped-query attention Turning pointLlama 3.1 405B built directly on thisChapter 9. Making it run anywhere
  8. 2023LLaMA Turning pointPosition interpolation built directly on thisChapter 8. The weights get out
  9. 2022Chinchilla Turning pointLlama 3.1 405B was influenced by thisChapter 5. Scale becomes a science
  10. 2022InstructGPT Turning pointDPO built directly on thisChapter 6. Teaching models what people want
  11. 2021GopherChinchilla built directly on thisChapter 5. Scale becomes a science
  12. 2021FLANInstructGPT was influenced by thisChapter 6. Teaching models what people want
  13. 2021RoPE Turning pointPosition interpolation built directly on thisChapter 10. Reading more than you remember
  14. 2021Switch TransformerLlama 4 was influenced by thisChapter 5. Scale becomes a science
  15. 2020RLHF for summarisation Turning pointDPO built directly on thisChapter 6. Teaching models what people want
  16. 2020GShard Turning pointSwitch Transformer built directly on thisChapter 5. Scale becomes a science
  17. 2020GPT-3 Turning pointInstructGPT built directly on thisChapter 5. Scale becomes a science
  18. 2020Scaling laws Turning pointChinchilla built directly on thisChapter 5. Scale becomes a science
  19. 2019Multi-query attentionGrouped-query attention built directly on thisChapter 9. Making it run anywhere
  20. 2019T5 Turning pointSwitch Transformer built directly on thisChapter 4. Pretrain once, use everywhere
  21. 2019C4 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  22. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  23. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  24. 2019GPT-2 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  25. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  26. 2019Transformer-XLRoPE was influenced by thisChapter 4. Pretrain once, use everywhere
  27. 2018BERT Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  28. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  29. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  30. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  31. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  32. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  33. 2017Transformer Turning pointT5 built directly on thisChapter 3. Attention, then only attention
  34. 2017RLHF originRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  35. 2017Sparse MoE Turning pointSwitch Transformer built directly on thisChapter 3. Attention, then only attention
  36. 2016Layer normTransformer built directly on this
    GPT-2 built directly on this    Chapter 3. Attention, then only attention
  37. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  38. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  39. 2015BPE tokenizationGPT-2 built directly on this
    LLaMA built directly on this    Chapter 2. Words become vectors
  40. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  41. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  42. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  43. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  44. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  45. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  46. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  47. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  48. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  49. 1997LSTM Turning pointSparse MoE built directly on thisChapter 1. Language as probability
  50. 1991Mixture of expertsSparse MoE built directly on thisChapter 1. Language as probability
  51. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2025Llama 4This milestone.
  2. 2024Llama 3.1 405BLlama 4 succeeded thisChapter 8. The weights get out
  3. 2024Llama 3Llama 3.1 405B built directly on thisChapter 8. The weights get out
  4. 2023Mixtral 8x7BLlama 4 was influenced by this (indirect influence)Chapter 8. The weights get out
  5. 2023Mistral 7BMixtral 8x7B built directly on thisChapter 8. The weights get out
  6. 2023Llama 2 Turning pointLlama 3 built directly on thisChapter 8. The weights get out
  7. 2023Position interpolationLlama 3.1 405B was influenced by thisChapter 10. Reading more than you remember
  8. 2023DPOLlama 3.1 405B built directly on thisChapter 6. Teaching models what people want
  9. 2023Grouped-query attention Turning pointLlama 3.1 405B built directly on thisChapter 9. Making it run anywhere
  10. 2023LLaMA Turning pointPosition interpolation built directly on thisChapter 8. The weights get out
  11. 2022FlashAttentionMistral 7B was influenced by thisChapter 9. Making it run anywhere
  12. 2022Chinchilla Turning pointLlama 3.1 405B was influenced by thisChapter 5. Scale becomes a science
  13. 2022InstructGPT Turning pointDPO built directly on thisChapter 6. Teaching models what people want
  14. 2021GopherChinchilla built directly on thisChapter 5. Scale becomes a science
  15. 2021FLANInstructGPT was influenced by thisChapter 6. Teaching models what people want
  16. 2021RoPE Turning pointPosition interpolation built directly on thisChapter 10. Reading more than you remember
  17. 2021Switch TransformerLlama 4 was influenced by thisChapter 5. Scale becomes a science
  18. 2020RLHF for summarisation Turning pointDPO built directly on thisChapter 6. Teaching models what people want
  19. 2020BigBirdFlashAttention succeeded thisChapter 10. Reading more than you remember
  20. 2020GShard Turning pointSwitch Transformer built directly on this
    Mixtral 8x7B built directly on this    Chapter 5. Scale becomes a science
  21. 2020GPT-3 Turning pointInstructGPT built directly on thisChapter 5. Scale becomes a science
  22. 2020LongformerMistral 7B was influenced by thisChapter 10. Reading more than you remember
  23. 2020Scaling laws Turning pointChinchilla built directly on thisChapter 5. Scale becomes a science
  24. 2019Multi-query attentionGrouped-query attention built directly on thisChapter 9. Making it run anywhere
  25. 2019T5 Turning pointSwitch Transformer built directly on thisChapter 4. Pretrain once, use everywhere
  26. 2019C4 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  27. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  28. 2019Sparse Transformer Turning pointMistral 7B was influenced by thisChapter 4. Pretrain once, use everywhere
  29. 2019GPT-2 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  30. 2019WebTextGPT-2 was influenced by this
    C4 was influenced by this (indirect influence)    Chapter 4. Pretrain once, use everywhere
  31. 2019Transformer-XLSparse Transformer was influenced by this (indirect influence)
    RoPE was influenced by this    Chapter 4. Pretrain once, use everywhere
  32. 2018BERT Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  33. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  34. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  35. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  36. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  37. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  38. 2017Transformer Turning pointT5 built directly on thisChapter 3. Attention, then only attention
  39. 2017RLHF originRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  40. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  41. 2017Sparse MoE Turning pointSwitch Transformer built directly on this
    Mixtral 8x7B was influenced by this    Chapter 3. Attention, then only attention
  42. 2016Layer normTransformer built directly on this
    GPT-2 built directly on this    Chapter 3. Attention, then only attention
  43. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  44. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  45. 2015BPE tokenizationGPT-2 built directly on this
    LLaMA built directly on this    Chapter 2. Words become vectors
  46. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  47. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  48. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  49. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  50. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  51. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  52. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  53. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  54. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  55. 1997LSTM Turning pointSparse MoE built directly on thisChapter 1. Language as probability
  56. 1991Mixture of expertsSparse MoE built directly on thisChapter 1. Language as probability
  57. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

Llama 4 has 1 documented descendants across 1 chapter, reaching forward to Muse Spark in 2026. Showing all of them.

Llama 4 has 1 documented descendants across 1 chapter, reaching forward to Muse Spark in 2026. Showing all of them.

Llama 4 has 1 recorded descendants, including indirect ones, across 1 chapter, reaching forward to Muse Spark in 2026. Showing all of them.

Llama 4 has 1 recorded descendants, including indirect ones, across 1 chapter, reaching forward to Muse Spark in 2026. Showing all of them.

  1. 2025Llama 4This milestone.
  2. 2026Muse Sparksucceeded Llama 4Chapter 8. The weights get out
  1. 2025Llama 4This milestone.
  2. 2026Muse Sparksucceeded Llama 4Chapter 8. The weights get out

Direct links and their sources

Where it came from

Influenced by

Succeeded

What it led to

Replaced by