By parameters
Six steps, from the smallest to the largest. Each model is placed by its parameters in billion parameters, where its sources give one.
Small
under 1 billion. BERT, GPT-2 small and the tiny models that run on a phone or a laptop.
6 models
Compact
1 to 10 billion. Phi, Gemma, Llama 3 8B, Mistral 7B and most on-device and edge models.
82 models
Mid-size
10 to 70 billion. The workhorse open models: Llama 2 13B, Qwen 32B, Gemma 27B and their kin.
103 models
Very large
300 billion to 1 trillion. PaLM, Llama 3.1 405B, DeepSeek-V3 and the big mixture-of-experts models.
37 models
Trillion-scale
1 trillion and more. Kimi K2, the Switch Transformer and the largest models whose size is published.
19 models