列名的含义
- M: 模型参数量
- T: 模型预训练tokens
- zh: 模型是否很好的支持中文
- code: 模型是否支持code
- S: 模型的context sequence长度, 1K=1024
- L: 模型non-embedding的层数
- D: 模型token embedding维度
- F: FFN中隐向量dff维度
- H: Multi-head attention中head数
- G: Grouped Query Attention (GQA)中key/value head数
- V: 词表大小
- Type: 模型类型(Dense/MoE/Mamba/...)
- Tie: 是否tie embedding
| LLM | M | T | zh | code | S | L | D | F | H | G | V | Tie | Type | 时间 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-1 | 117M | 800M(w) | False | False | 512 | 12 | 768 | 3072 | 12 | 12 | 40478 | True | Dense | 2018-06-11 |
| GPT-2 | 124M | 10B | False | False | 1K | 12 | 768 | 3072 | 12 | 12 | 50257 | True | Dense | 2019-02-14 |
| 355M | 10B | False | False | 1K | 24 | 1024 | 4096 | 16 | 16 | 50257 | True | Dense | 2019-02-14 | |
| 774M | 10B | False | False | 1K | 36 | 1280 | 5120 | 20 | 20 | 50257 | True | Dense | 2019-02-14 | |
| 1.5B | 10B | False | False | 1K | 48 | 1600 | 6400 | 25 | 25 | 50257 | True | Dense | 2019-02-14 | |
| GPT-3 | 125M | 300B | False | False | 2K | 12 | 768 | 3072 | 12 | 12 | 50257 | True | Dense | 2020-05-28 |
| 350M | 300B | False | False | 2K | 24 | 1024 | 4096 | 16 | 16 | 50257 | ?True | Dense | 2020-05-28 | |
| 760M | 300B | False | False | 2K | 24 | 1536 | 6144 | 16 | 16 | 50257 | ?True | Dense | 2020-05-28 | |
| 1.3B | 300B | False | False | 2K | 24 | 2048 | 8192 | 24 | 24 | 50257 | ?True | Dense | 2020-05-28 | |
| 2.7B | 300B | False | False | 2K | 32 | 2560 | 10240 | 32 | 32 | 50257 | ?True | Dense | 2020-05-28 | |
| 6.7B | 300B | False | False | 2K | 32 | 4096 | 16384 | 32 | 32 | 50257 | ?True | Dense | 2020-05-28 | |
| 13B | 300B | False | False | 2K | 40 | 5120 | 20480 | 40 | 40 | 50257 | ?True | Dense | 2020-05-28 | |
| 175B | 300B | False | False | 2K | 96 | 12288 | 49152 | 96 | 96 | 50257 | ?True | Dense | 2020-05-28 | |
| Llama 1 | 7B | 1T | False | False | 2K | 32 | 4096 | 11008 | 32 | 32 | 32000 | False | Dense | 2023-02-24 |
| 13B | 1T | False | True | 2K | 40 | 5120 | 13824 | 40 | 40 | 32000 | False | Dense | 2023-02-24 | |
| 33B | 1.4T | False | True | 2K | 60 | 6656 | 17920 | 52 | 52 | 32000 | False | Dense | 2023-02-24 | |
| 65B | 1.4T | False | True | 2K | 80 | 8192 | 22016 | 64 | 64 | 32000 | False | Dense | 2023-02-24 | |
| Llama 2 | 7B | 2T | False | True | 4K | 32 | 4096 | 11008 | 32 | 32 | 32000 | False | Dense | 2023-07-18 |
| 13B | 2T | False | True | 4K | 40 | 5120 | 13824 | 40 | 40 | 32000 | False | Dense | 2023-07-18 | |
| 70B | 2T | False | True | 4K | 80 | 8192 | 28672 | 64 | 8 | 32000 | False | Dense | 2023-07-18 | |
| Llama 3 | 8B | 15T | False | True | 8K | 32 | 4096 | 14336 | 32 | 8 | 128256 | False | Dense | 2024-04-18 |
| 70B | 15T | False | True | 8K | 80 | 8192 | 28672 | 64 | 8 | 128256 | False | Dense | 2024-04-18 | |
| Llama 3.1 | 8B | 15T | False | True | 128K | 32 | 4096 | 14336 | 32 | 8 | 128256 | False | Dense | 2024-07-23 |
| 70B | 15T | False | True | 128K | 80 | 8192 | 28672 | 64 | 8 | 128256 | False | Dense | 2024-07-23 | |
| 405B | 15T | False | True | 128K | 126 | 16384 | 53248 | 128 | 8 | 128256 | False | Dense | 2024-07-23 | |
| Llama 3.2 | 1B | 15T | False | True | 128K | 16 | 2K | 8192 | 32 | 8 | 128256 | True | Dense | 2024-09-25 |
| 3B | 15T | False | True | 128K | 28 | 3072 | 8192 | 24 | 8 | 128256 | True | Dense | 2024-09-25 | |
| Llama 3.3 | 70B | 15T | False | True | 128K | 80 | 8192 | 28672 | 64 | 8 | 128256 | False | Dense | 2024-12-06 |
| Qwen 1 | 1.8B | 2.2T | True | True | 8K | 24 | 2048 | 11008 | 16 | 16 | 151936 | False | Dense | 2023-11-30 |
| 7B | 2.4T | True | True | 8K | 32 | 4096 | 22016 | 32 | 32 | 151936 | False | Dense | 2023-11-30 | |
| 14B | 3T | True | True | 2K | 40 | 5120 | 27392 | 40 | 40 | 152064 | False | Dense | 2023-11-30 | |
| 72B | 3T | True | True | 32K | 80 | 8192 | 49152 | 64 | 64 | 152064 | False | Dense | 2023-11-30 | |
| Qwen1.5 | 0.5B | 2.4T | True | True | 32K | 24 | 1024 | 2816 | 16 | 16 | 151936 | True | Dense | 2024-02-04 |
| 1.8B | 2.4T | True | True | 32K | 24 | 2048 | 5504 | 16 | 16 | 151936 | False | Dense | 2024-02-04 | |
| 4B | 2.4T | True | True | 32K | 40 | 2560 | 6912 | 20 | 20 | 151936 | False | Dense | 2024-02-04 | |
| 7B | 4T | True | True | 32K | 32 | 4096 | 11008 | 32 | 32 | 151936 | False | Dense | 2024-02-04 | |
| 14B | 4T | True | True | 32K | 40 | 5120 | 13696 | 40 | 40 | 152064 | False | Dense | 2024-02-04 | |
| 32B | ?T | True | True | 32K | 64 | 5120 | 27392 | 40 | 8 | 152064 | False | Dense | 2024-02-04 | |
| 72B | 3T | True | True | 32K | 80 | 8192 | 24576 | 64 | 64 | 152064 | False | Dense | 2024-02-04 | |
| 110B | ?T | True | True | 32K | 80 | 8192 | 49152 | 64 | 8 | 152064 | False | Dense | 2024-02-04 | |
| MoE-A2.7B | ?T | True | True | 32K | 24 | 2048 | 5632 | 16 | 16 | 151936 | False | MoE | 2024-02-04 | |
| Qwen2 | 0.5B | 12T | True | True | 128K | 24 | 896 | 4864 | 14 | 2 | 151646 | True | Dense | 2024-06-07 |
| 1.5B | 7T | True | True | 128K | 28 | 1536 | 8960 | 12 | 2 | 151646 | True | Dense | 2024-06-07 | |
| 7B | 7T | True | True | 128K | 28 | 3584 | 18944 | 28 | 4 | 151646 | False | Dense | 2024-06-07 | |
| 72B | 7T | True | True | 128K | 80 | 8192 | 29568 | 64 | 8 | 151646 | False | Dense | 2024-06-07 | |
| 57B-A14B | 4.5T | True | True | 128K | 28 | 3584 | 2560 | 28 | 4 | 151646 | False | MoE | 2024-06-07 | |
| Qwen2.5 | 0.5B | ?18T | True | True | 32K | 24 | 896 | 4864 | 14 | 2 | 151936 | True | Dense | 2024-09-15 |
| 1.5B | ?18T | True | True | 32K | 28 | 1536 | 8960 | 12 | 2 | 151936 | True | Dense | 2024-09-15 | |
| 3B | ?18T | True | True | 32K | 36 | 2048 | 11008 | 16 | 2 | 151936 | True | Dense | 2024-09-15 | |
| 7B | ?18T | True | True | 128K | 28 | 3584 | 18944 | 28 | 4 | 151936 | False | Dense | 2024-09-15 | |
| 14B | ?18T | True | True | 128K | 48 | 5120 | 13824 | 40 | 8 | 151936 | False | Dense | 2024-09-15 | |
| 32B | ?18T | True | True | 128K | 64 | 5120 | 27648 | 40 | 8 | 152064 | False | Dense | 2024-09-15 | |
| 72B | ?18T | True | True | 128K | 80 | 8192 | 29568 | 64 | 8 | 152064 | False | Dense | 2024-09-15 | |
| DeepSeek LLM | 7B | 2T | True | True | 4K | 30 | 4096 | 11008 | 32 | 32 | 102400 | False | Dense | 2023-11-29 |
| 67B | 2T | True | True | 4K | 95 | 8192 | 22016 | 64 | 8 | 102400 | False | Dense | 2023-11-29 | |
| DeepSeekMoE | 16B | 2T | True | True | 4K | 28 | 2048 | 10944 | 16 | 16 | 102400 | False | MoE | 2024-01-08 |