NLP Lineage

11 July 2025 · Language modelling

Kimi K2

Kimi K2

Kimi Team

Date
11 July 2025
What the date marks
Model release
Kind
Model
Organisation
Moonshot AI
Access
modified MIT open weights
Other dates
arXiv 2025-07-28
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

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

Open MoE models topped out well below a trillion parameters.

The problem

Training at that scale is unstable, and agentic skill needs data nobody had assembled.

What it introduced

A roughly one-trillion-parameter open MoE with about 32B active, trained with the MuonClip optimizer and heavy agentic post-training.

Why it matters

Open weights reached trillion-parameter scale with agentic capability, from outside the established labs.

What modern models inherit

Trillion-parameter open models as a real category.

What came next

Kimi K3 in 2026 took it to multimodal at a larger scale.

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

Kimi K2 has 49 documented ancestors across 10 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 21 turning points among them.

Kimi K2 has 49 documented ancestors across 10 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

Kimi K2 has 67 recorded ancestors, including indirect ones, across 12 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 24 turning points among them.

Kimi K2 has 67 recorded ancestors, including indirect ones, across 12 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2025Kimi K2This milestone.
  2. 2024GRPOKimi K2 was influenced by thisChapter 12. Thinking longer
  3. 2024DeepSeekMoEKimi K2 was influenced by thisChapter 8. The weights get out
  4. 2024DeepSeek LLMGRPO built directly on thisChapter 8. The weights get out
  5. 2023Llama 2 Turning pointDeepSeek LLM was influenced by thisChapter 8. The weights get out
  6. 2023Grouped-query attention Turning pointLlama 2 built directly on thisChapter 9. Making it run anywhere
  7. 2023LLaMA Turning pointLlama 2 built directly on thisChapter 8. The weights get out
  8. 2022Chinchilla Turning pointDeepSeek LLM was influenced by thisChapter 5. Scale becomes a science
  9. 2022InstructGPT Turning pointGRPO was influenced by thisChapter 6. Teaching models what people want
  10. 2021GopherChinchilla built directly on thisChapter 5. Scale becomes a science
  11. 2021FLANInstructGPT was influenced by thisChapter 6. Teaching models what people want
  12. 2021RoPE Turning pointLLaMA built directly on thisChapter 10. Reading more than you remember
  13. 2021Switch TransformerDeepSeekMoE was influenced by thisChapter 5. Scale becomes a science
  14. 2020RLHF for summarisation Turning pointInstructGPT built directly on thisChapter 6. Teaching models what people want
  15. 2020GShard Turning pointDeepSeekMoE built directly on thisChapter 5. Scale becomes a science
  16. 2020GPT-3 Turning pointInstructGPT built directly on thisChapter 5. Scale becomes a science
  17. 2020Scaling laws Turning pointDeepSeek LLM built directly on thisChapter 5. Scale becomes a science
  18. 2019Multi-query attentionGrouped-query attention built directly on thisChapter 9. Making it run anywhere
  19. 2019T5 Turning pointSwitch Transformer built directly on thisChapter 4. Pretrain once, use everywhere
  20. 2019C4 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  21. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  22. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  23. 2019GPT-2 Turning pointT5 was influenced by this
    Scaling laws was influenced by this
    GPT-3 built directly on this    Chapter 4. Pretrain once, use everywhere
  24. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  25. 2019Transformer-XLRoPE was influenced by thisChapter 4. Pretrain once, use everywhere
  26. 2018BERT Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  27. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  28. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  29. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  30. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  31. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  32. 2017Transformer Turning pointT5 built directly on this
    Scaling laws built directly on this    Chapter 3. Attention, then only attention
  33. 2017RLHF originRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  34. 2017Sparse MoE Turning pointDeepSeekMoE was influenced by thisChapter 3. Attention, then only attention
  35. 2016Layer normTransformer built directly on this
    GPT-2 built directly on this    Chapter 3. Attention, then only attention
  36. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  37. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  38. 2015BPE tokenizationGPT-2 built directly on this
    LLaMA built directly on this    Chapter 2. Words become vectors
  39. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  40. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  41. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  42. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  43. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  44. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  45. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  46. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  47. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  48. 1997LSTM Turning pointSparse MoE built directly on thisChapter 1. Language as probability
  49. 1991Mixture of expertsSparse MoE built directly on thisChapter 1. Language as probability
  50. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2025Kimi K2This milestone.
  2. 2024DeepSeek-V3Kimi K2 was influenced by this (indirect influence)Chapter 8. The weights get out
  3. 2024Llama 3.1 405BDeepSeek-V3 was influenced by thisChapter 8. The weights get out
  4. 2024GPT-4oDeepSeek-V3 was influenced by thisChapter 11. Seeing and hearing
  5. 2024DeepSeek-V2 / MLADeepSeek-V3 built directly on thisChapter 9. Making it run anywhere
  6. 2024Llama 3Llama 3.1 405B built directly on thisChapter 8. The weights get out
  7. 2024GRPOKimi K2 was influenced by thisChapter 12. Thinking longer
  8. 2024DeepSeekMoEKimi K2 was influenced by thisChapter 8. The weights get out
  9. 2024DeepSeek LLMGRPO built directly on thisChapter 8. The weights get out
  10. 2023GPT-4VGPT-4o built directly on thisChapter 11. Seeing and hearing
  11. 2023Llama 2 Turning pointDeepSeek LLM was influenced by thisChapter 8. The weights get out
  12. 2023Position interpolationLlama 3.1 405B was influenced by thisChapter 10. Reading more than you remember
  13. 2023Process supervisionGRPO was influenced by this (indirect influence)Chapter 12. Thinking longer
  14. 2023DPOLlama 3.1 405B built directly on thisChapter 6. Teaching models what people want
  15. 2023Grouped-query attention Turning pointDeepSeek-V2 / MLA was influenced by this
    Llama 3.1 405B built directly on this    Chapter 9. Making it run anywhere
  16. 2023GPT-4 Turning pointGPT-4o built directly on thisChapter 7. The assistant goes public
  17. 2023LLaMA Turning pointLlama 2 built directly on this
    Position interpolation built directly on this    Chapter 8. The weights get out
  18. 2022GPT-3.5 Turning pointGPT-4 succeeded thisChapter 7. The assistant goes public
  19. 2022WhisperGPT-4o was influenced by this (indirect influence)Chapter 11. Seeing and hearing
  20. 2022Chinchilla Turning pointDeepSeek LLM was influenced by this
    Llama 3.1 405B was influenced by this    Chapter 5. Scale becomes a science
  21. 2022STaRProcess supervision was influenced by this (indirect influence)Chapter 12. Thinking longer
  22. 2022InstructGPT Turning pointGRPO was influenced by thisChapter 6. Teaching models what people want
  23. 2022Chain of thought Turning pointProcess supervision built directly on thisChapter 12. Thinking longer
  24. 2021GopherChinchilla built directly on thisChapter 5. Scale becomes a science
  25. 2021FLANInstructGPT was influenced by thisChapter 6. Teaching models what people want
  26. 2021RoPE Turning pointDeepSeek-V2 / MLA built directly on thisChapter 10. Reading more than you remember
  27. 2021CLIPGPT-4V was influenced by this (indirect influence)Chapter 11. Seeing and hearing
  28. 2021Switch TransformerDeepSeekMoE was influenced by thisChapter 5. Scale becomes a science
  29. 2020Vision TransformerCLIP built directly on thisChapter 11. Seeing and hearing
  30. 2020RLHF for summarisation Turning pointInstructGPT built directly on thisChapter 6. Teaching models what people want
  31. 2020GShard Turning pointDeepSeekMoE built directly on thisChapter 5. Scale becomes a science
  32. 2020GPT-3 Turning pointInstructGPT built directly on thisChapter 5. Scale becomes a science
  33. 2020Scaling laws Turning pointDeepSeek LLM built directly on thisChapter 5. Scale becomes a science
  34. 2019Multi-query attentionDeepSeek-V2 / MLA was influenced by thisChapter 9. Making it run anywhere
  35. 2019T5 Turning pointSwitch Transformer built directly on thisChapter 4. Pretrain once, use everywhere
  36. 2019C4 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  37. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  38. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  39. 2019GPT-2 Turning pointT5 was influenced by this
    Scaling laws was influenced by this
    GPT-3 built directly on this    Chapter 4. Pretrain once, use everywhere
  40. 2019WebTextGPT-2 was influenced by this
    C4 was influenced by this (indirect influence)    Chapter 4. Pretrain once, use everywhere
  41. 2019Transformer-XLRoPE was influenced by thisChapter 4. Pretrain once, use everywhere
  42. 2018BERT Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  43. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  44. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  45. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  46. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  47. 2017Mixed precisionDeepSeek-V3 was influenced by thisChapter 9. Making it run anywhere
  48. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  49. 2017Transformer Turning pointT5 built directly on this
    Scaling laws built directly on this
    RoPE built directly on this    Chapter 3. Attention, then only attention
  50. 2017RLHF originRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  51. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  52. 2017Sparse MoE Turning pointDeepSeekMoE was influenced by thisChapter 3. Attention, then only attention
  53. 2016Layer normTransformer built directly on this
    GPT-2 built directly on this    Chapter 3. Attention, then only attention
  54. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  55. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  56. 2015BPE tokenizationGPT-2 built directly on this
    LLaMA built directly on this    Chapter 2. Words become vectors
  57. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  58. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  59. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  60. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  61. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  62. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  63. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  64. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  65. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  66. 1997LSTM Turning pointSparse MoE built directly on thisChapter 1. Language as probability
  67. 1991Mixture of expertsSparse MoE built directly on thisChapter 1. Language as probability
  68. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

Kimi K2 has 1 documented descendants across 1 chapter, reaching forward to Kimi K3 in 2026. Showing all of them.

Kimi K2 has 1 documented descendants across 1 chapter, reaching forward to Kimi K3 in 2026. Showing all of them.

Kimi K2 has 1 recorded descendants, including indirect ones, across 1 chapter, reaching forward to Kimi K3 in 2026. Showing all of them.

Kimi K2 has 1 recorded descendants, including indirect ones, across 1 chapter, reaching forward to Kimi K3 in 2026. Showing all of them.

  1. 2025Kimi K2This milestone.
  2. 2026Kimi K3built directly on Kimi K2Chapter 8. The weights get out
  1. 2025Kimi K2This milestone.
  2. 2026Kimi K3built directly on Kimi K2Chapter 8. The weights get out

Direct links and their sources

Where it came from

Influenced by

What it led to

Built on directly by