List of large language models
From Wikipedia, the free encyclopedia
A large language model (LLM) is a type of machine learning model designed for natural language processing tasks such as language generation. LLMs are language models with many parameters, and are trained with self-supervised learning on a vast amount of text.
This page lists notable large language models.
List
Summarize
Perspective
For the training cost column, 1 petaFLOP-day = 1 petaFLOP/sec × 1 day = 8.64E19 FLOP. Also, only the largest model's cost is written.
Name | Release date[a] | Developer | Number of parameters (billion) [b] | Corpus size | Training cost (petaFLOP-day) | License[c] | Notes |
---|---|---|---|---|---|---|---|
Attention Is All You Need | June 2017 | Vaswani et al at Google | 0.213 | 36 million English-French sentence pairs | 0.09[1] | Trained for 0.3M steps on 8 NVIDIA P100 GPUs. | |
GPT-1 | June 2018 | OpenAI | 0.117 | 1[2] | MIT[3] | First GPT model, decoder-only transformer. Trained for 30 days on 8 P600 GPUs. | |
BERT | October 2018 | 0.340[4] | 3.3 billion words[4] | 9[5] | Apache 2.0[6] | An early and influential language model.[7]Encoder-only and thus not built to be prompted or generative.[8] Training took 4 days on 64 TPUv2 chips.[9] | |
T5 | October 2019 | 11[10] | 34 billion tokens[10] | Apache 2.0[11] | Base model for many Google projects, such as Imagen.[12] | ||
XLNet | June 2019 | 0.340[13] | 33 billion words | 330 | Apache 2.0[14] | An alternative to BERT; designed as encoder-only. Trained on 512 TPU v3 chips for 5.5 days.[15] | |
GPT-2 | February 2019 | OpenAI | 1.5[16] | 40GB[17] (~10 billion tokens)[18] | 28[19] | MIT[20] | Trained on 32 TPUv3 chips for 1 week.[19] |
GPT-3 | May 2020 | OpenAI | 175[21] | 300 billion tokens[18] | 3640[22] | Proprietary | A fine-tuned variant of GPT-3, termed GPT-3.5, was made available to the public through a web interface called ChatGPT in 2022.[23] |
GPT-Neo | March 2021 | EleutherAI | 2.7[24] | 825 GiB[25] | MIT[26] | The first of a series of free GPT-3 alternatives released by EleutherAI. GPT-Neo outperformed an equivalent-size GPT-3 model on some benchmarks, but was significantly worse than the largest GPT-3.[26] | |
GPT-J | June 2021 | EleutherAI | 6[27] | 825 GiB[25] | 200[28] | Apache 2.0 | GPT-3-style language model |
Megatron-Turing NLG | October 2021 [29] | Microsoft and Nvidia | 530[30] | 338.6 billion tokens[30] | 38000[31] | Restricted web access | Trained for 3 months on over 2000 A100 GPUs on the NVIDIA Selene Supercomputer, for over 3 million GPU-hours.[31] |
Ernie 3.0 Titan | December 2021 | Baidu | 260[32] | 4 Tb | Proprietary | Chinese-language LLM. Ernie Bot is based on this model. | |
Claude[33] | December 2021 | Anthropic | 52[34] | 400 billion tokens[34] | beta | Fine-tuned for desirable behavior in conversations.[35] | |
GLaM (Generalist Language Model) | December 2021 | 1200[36] | 1.6 trillion tokens[36] | 5600[36] | Proprietary | Sparse mixture of experts model, making it more expensive to train but cheaper to run inference compared to GPT-3. | |
Gopher | December 2021 | DeepMind | 280[37] | 300 billion tokens[38] | 5833[39] | Proprietary | Later developed into the Chinchilla model. |
LaMDA (Language Models for Dialog Applications) | January 2022 | 137[40] | 1.56T words,[40] 168 billion tokens[38] | 4110[41] | Proprietary | Specialized for response generation in conversations. | |
GPT-NeoX | February 2022 | EleutherAI | 20[42] | 825 GiB[25] | 740[28] | Apache 2.0 | based on the Megatron architecture |
Chinchilla | March 2022 | DeepMind | 70[43] | 1.4 trillion tokens[43][38] | 6805[39] | Proprietary | Reduced-parameter model trained on more data. Used in the Sparrow bot. Often cited for its neural scaling law. |
PaLM (Pathways Language Model) | April 2022 | 540[44] | 768 billion tokens[43] | 29,250[39] | Proprietary | Trained for ~60 days on ~6000 TPU v4 chips.[39] As of October 2024[update], it is the largest dense Transformer published. | |
OPT (Open Pretrained Transformer) | May 2022 | Meta | 175[45] | 180 billion tokens[46] | 310[28] | Non-commercial research[d] | GPT-3 architecture with some adaptations from Megatron. Uniquely, the training logbook written by the team was published.[47] |
YaLM 100B | June 2022 | Yandex | 100[48] | 1.7TB[48] | Apache 2.0 | English-Russian model based on Microsoft's Megatron-LM. | |
Minerva | June 2022 | 540[49] | 38.5B tokens from webpages filtered for mathematical content and from papers submitted to the arXiv preprint server[49] | Proprietary | For solving "mathematical and scientific questions using step-by-step reasoning".[50] Initialized from PaLM models, then finetuned on mathematical and scientific data. | ||
BLOOM | July 2022 | Large collaboration led by Hugging Face | 175[51] | 350 billion tokens (1.6TB)[52] | Responsible AI | Essentially GPT-3 but trained on a multi-lingual corpus (30% English excluding programming languages) | |
Galactica | November 2022 | Meta | 120 | 106 billion tokens[53] | Unknown | CC-BY-NC-4.0 | Trained on scientific text and modalities. |
AlexaTM (Teacher Models) | November 2022 | Amazon | 20[54] | 1.3 trillion[55] | Proprietary[56] | bidirectional sequence-to-sequence architecture | |
LLaMA (Large Language Model Meta AI) | February 2023 | Meta AI | 65[57] | 1.4 trillion[57] | 6300[58] | Non-commercial research[e] | Corpus has 20 languages. "Overtrained" (compared to Chinchilla scaling law) for better performance with fewer parameters.[57] |
GPT-4 | March 2023 | OpenAI | Unknown[f] (According to rumors: 1760)[60] |
Unknown | Unknown, estimated 230,000. | Proprietary | Available for ChatGPT Plus users and used in several products. |
Chameleon | June 2024 | Meta AI | 34[61] | 4.4 trillion | |||
Cerebras-GPT | March 2023 | Cerebras | 13[62] | 270[28] | Apache 2.0 | Trained with Chinchilla formula. | |
Falcon | March 2023 | Technology Innovation Institute | 40[63] | 1 trillion tokens, from RefinedWeb (filtered web text corpus)[64] plus some "curated corpora".[65] | 2800[58] | Apache 2.0[66] | |
BloombergGPT | March 2023 | Bloomberg L.P. | 50 | 363 billion token dataset based on Bloomberg's data sources, plus 345 billion tokens from general purpose datasets[67] | Proprietary | Trained on financial data from proprietary sources, for financial tasks. | |
PanGu-Σ | March 2023 | Huawei | 1085 | 329 billion tokens[68] | Proprietary | ||
OpenAssistant[69] | March 2023 | LAION | 17 | 1.5 trillion tokens | Apache 2.0 | Trained on crowdsourced open data | |
Jurassic-2[70] | March 2023 | AI21 Labs | Unknown | Unknown | Proprietary | Multilingual[71] | |
PaLM 2 (Pathways Language Model 2) | May 2023 | 340[72] | 3.6 trillion tokens[72] | 85,000[58] | Proprietary | Was used in Bard chatbot.[73] | |
Llama 2 | July 2023 | Meta AI | 70[74] | 2 trillion tokens[74] | 21,000 | Llama 2 license | 1.7 million A100-hours.[75] |
Claude 2 | July 2023 | Anthropic | Unknown | Unknown | Unknown | Proprietary | Used in Claude chatbot.[76] |
Granite 13b | July 2023 | IBM | Unknown | Unknown | Unknown | Proprietary | Used in IBM Watsonx.[77] |
Mistral 7B | September 2023 | Mistral AI | 7.3[78] | Unknown | Apache 2.0 | ||
Claude 2.1 | November 2023 | Anthropic | Unknown | Unknown | Unknown | Proprietary | Used in Claude chatbot. Has a context window of 200,000 tokens, or ~500 pages.[79] |
Grok 1[80] | November 2023 | xAI | 314 | Unknown | Unknown | Apache 2.0 | Used in Grok chatbot. Grok 1 has a context length of 8,192 tokens and has access to X (Twitter).[81] |
Gemini 1.0 | December 2023 | Google DeepMind | Unknown | Unknown | Unknown | Proprietary | Multimodal model, comes in three sizes. Used in the chatbot of the same name.[82] |
Mixtral 8x7B | December 2023 | Mistral AI | 46.7 | Unknown | Unknown | Apache 2.0 | Outperforms GPT-3.5 and Llama 2 70B on many benchmarks.[83] Mixture of experts model, with 12.9 billion parameters activated per token.[84] |
Mixtral 8x22B | April 2024 | Mistral AI | 141 | Unknown | Unknown | Apache 2.0 | [85] |
DeepSeek-LLM | November 29, 2023 | DeepSeek | 67 | 2T tokens[86]: table 2 | 12,000 | DeepSeek License | Trained on English and Chinese text. 1e24 FLOPs for 67B. 1e23 FLOPs for 7B[86]: figure 5 |
Phi-2 | December 2023 | Microsoft | 2.7 | 1.4T tokens | 419[87] | MIT | Trained on real and synthetic "textbook-quality" data, for 14 days on 96 A100 GPUs.[87] |
Gemini 1.5 | February 2024 | Google DeepMind | Unknown | Unknown | Unknown | Proprietary | Multimodal model, based on a Mixture-of-Experts (MoE) architecture. Context window above 1 million tokens.[88] |
Gemini Ultra | February 2024 | Google DeepMind | Unknown | Unknown | Unknown | ||
Gemma | February 2024 | Google DeepMind | 7 | 6T tokens | Unknown | Gemma Terms of Use[89] | |
Claude 3 | March 2024 | Anthropic | Unknown | Unknown | Unknown | Proprietary | Includes three models, Haiku, Sonnet, and Opus.[90] |
Nova | October 2024 | Rubik's AI | Unknown | Unknown | Unknown | Proprietary | Previous three models, Nova-Instant, Nova-Air, and Nova-Pro. Company shifted to Sonus AI. |
Sonus[91] | January 2025 | Rubik's AI | Unknown | Unknown | Unknown | Proprietary | |
DBRX | March 2024 | Databricks and Mosaic ML | 136 | 12T Tokens | Databricks Open Model License | Training cost 10 million USD. | |
Fugaku-LLM | May 2024 | Fujitsu, Tokyo Institute of Technology, etc. | 13 | 380B Tokens | The largest model ever trained on CPU-only, on the Fugaku.[92] | ||
Phi-3 | April 2024 | Microsoft | 14[93] | 4.8T Tokens | MIT | Microsoft markets them as "small language model".[94] | |
Granite Code Models | May 2024 | IBM | Unknown | Unknown | Unknown | Apache 2.0 | |
Qwen2 | June 2024 | Alibaba Cloud | 72[95] | 3T Tokens | Unknown | Qwen License | Multiple sizes, the smallest being 0.5B. |
DeepSeek-V2 | June 2024 | DeepSeek | 236 | 8.1T tokens | 28,000 | DeepSeek License | 1.4M hours on H800.[96] |
Nemotron-4 | June 2024 | Nvidia | 340 | 9T Tokens | 200,000 | NVIDIA Open Model License | Trained for 1 epoch. Trained on 6144 H100 GPUs between December 2023 and May 2024.[97][98] |
Llama 3.1 | July 2024 | Meta AI | 405 | 15.6T tokens | 440,000 | Llama 3 license | 405B version took 31 million hours on H100-80GB, at 3.8E25 FLOPs.[99][100] |
DeepSeek-V3 | December 2024 | DeepSeek | 671 | 14.8T tokens | 56,000 | MIT | 2.788M hours on H800 GPUs.[101] Originally released under the DeepSeek License, then re-released under the MIT License as "DeepSeek-V3-0324" in March 2025.[102] |
Amazon Nova | December 2024 | Amazon | Unknown | Unknown | Unknown | Proprietary | Includes three models, Nova Micro, Nova Lite, and Nova Pro[103] |
DeepSeek-R1 | January 2025 | DeepSeek | 671 | Not applicable | Unknown | MIT | No pretraining. Reinforcement-learned upon V3-Base.[104][105] |
Qwen2.5 | January 2025 | Alibaba | 72 | 18T tokens | Unknown | Qwen License | 7 dense models, with parameter count from 0.5B to 72B. They also released 2 MoE variants.[106] |
MiniMax-Text-01 | January 2025 | Minimax | 456 | 4.7T tokens[107] | Unknown | Minimax Model license | [108][107] |
Gemini 2.0 | February 2025 | Google DeepMind | Unknown | Unknown | Unknown | Proprietary | Three models released: Flash, Flash-Lite and Pro[109][110][111] |
Mistral Large | November 2024 | Mistral AI | 123 | Unknown | Unknown | Mistral Research License | Upgraded over time. The latest version is 24.11.[112] |
Pixtral | November 2024 | Mistral AI | 123 | Unknown | Unknown | Mistral Research License | Multimodal. There is also a 12B version which is under Apache 2 license.[112] |
Grok 3 | February 2025 | xAI | Unknown | Unknown | Unknown, estimated 5,800,000. |
Proprietary | Training cost claimed "10x the compute of previous state-of-the-art models".[113] |
Llama 4 | April 5, 2025 | Meta AI | 400 | 40T tokens | Llama 4 license | [114][115] | |
Qwen3 | April 2025 | Alibaba Cloud | 235 | 36T tokens | Unknown | Apache 2.0 | Multiple sizes, the smallest being 0.6B.[116] |
See also
Notes
- As stated in Technical report: "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method ..."[59]
References
Wikiwand - on
Seamless Wikipedia browsing. On steroids.