Models / Llama-2
Llama-2 by meta-llama
6 models (7B to 70B) · ↓ 1.1M downloads a month · from $0.10/h on RTX 3060 right now. Each model's versions do the same job; quantized ones (AWQ, FP8, 4-bit) need less GPU memory and run on cheaper machines.
| Version | Runtime | Weights | GPU memory | Cheapest machine now |
|---|
| Originalmeta-llama/Llama-2-7b-hfLicence to accept | vLLM | 13 GB | ~18.4 GB | $0.18/h 2x RTX 3060 |
| Version | Runtime | Weights | GPU memory | Cheapest machine now |
|---|
| Originalmeta-llama/Llama-2-7b-chat-hfLicence to accept | vLLM | 13 GB | ~18.4 GB | $0.18/h 2x RTX 3060 |
| Llama 2 7B Chat GPTQ (TheBloke)TheBloke/Llama-2-7B-Chat-GPTQ | vLLM | 3.9 GB | 8.9 GB | $0.10/h RTX 3060 |
| Version | Runtime | Weights | GPU memory | Cheapest machine now |
|---|
| Originalmeta-llama/Llama-2-13b-chat-hfLicence to accept | vLLM | 26 GB | ~33.5 GB | $0.45/h Nvidia A40 |
| Llama 2 13B chat GPTQ (TheBloke)TheBloke/Llama-2-13B-chat-GPTQ | vLLM | 7.3 GB | 14 GB | $0.14/h RTX A4000 |
| Version | Runtime | Weights | GPU memory | Cheapest machine now |
|---|
| Originalmeta-llama/Llama-2-13b-hfLicence to accept | vLLM | 26 GB | ~33.5 GB | $0.45/h Nvidia A40 |
| Llama 2 13B GPTQ (TheBloke)TheBloke/Llama-2-13B-GPTQ | vLLM | 7.3 GB | 14 GB | $0.14/h RTX A4000 |
| Version | Runtime | Weights | GPU memory | Cheapest machine now |
|---|
| Originalmeta-llama/Llama-2-70b-hfLicence to accept | vLLM | 138 GB | ~167.8 GB | $1.38/h 8x RTX 3090 |
| Llama 2 70B AWQ (TheBloke)TheBloke/Llama-2-70B-AWQ | vLLM | 37 GB | 44.4 GB | $0.45/h Nvidia A40 |
| Version | Runtime | Weights | GPU memory | Cheapest machine now |
|---|
| Originalmeta-llama/Llama-2-70b-chat-hfLicence to accept | vLLM | 138 GB | ~167.8 GB | $1.38/h 8x RTX 3090 |
| Llama 2 70B Chat GPTQ (TheBloke)TheBloke/Llama-2-70B-Chat-GPTQ | vLLM | 35 GB | 43 GB | $0.45/h Nvidia A40 |
Model details come from Hugging Face and are refreshed daily; prices follow the live offers. GPU memory is GridPup's estimate: the weights, room for an 8,192-token conversation (chat models) and the runtime's own memory. Large models are split across the GPUs of one machine ("4x RTX 4090").