Loading…
Loading…
Models / Llama-3.1
4 models (8B to 405B) · ↓ 8.6M downloads a month · from $0.09/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-3.1-8BLicence to accept | vLLM | 16 GB | ~21.5 GB | $0.18/h 2x RTX 3060 |
| Instructmeta-llama/Llama-3.1-8B-InstructLicence to accept | vLLM | 16 GB | ~21.5 GB | $0.18/h 2x RTX 3060 |
| Llama Guard 3 8Bmeta-llama/Llama-Guard-3-8BLicence to accept | vLLM | 16 GB | ~21.5 GB | $0.18/h 2x RTX 3060 |
| Instruct FP8 (RedHatAI)RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8 | vLLM | 9.1 GB | 13.5 GB | $0.14/h RTX A4000 |
| Instruct bnb 4bit (unsloth)unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit | vLLM | 5.7 GB | 9.8 GB | $0.10/h RTX 3060 |
| Instruct quantized.w4a16 (RedHatAI)RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w4a16 | vLLM | 5.7 GB | 9.8 GB | $0.10/h RTX 3060 |
| Instruct unsloth bnb 4bit (unsloth)unsloth/Meta-Llama-3.1-8B-Instruct-unsloth-bnb-4bit | vLLM | 6.0 GB | 10 GB | $0.10/h RTX 3060 |
| Instruct FP8 dynamic (RedHatAI)RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8-dynamic | vLLM | 9.1 GB | 13.5 GB | $0.14/h RTX A4000 |
| Instruct unsloth bnb 4bit (unsloth)unsloth/Llama-3.1-8B-Instruct-unsloth-bnb-4bit | vLLM | 6.0 GB | 10 GB | $0.10/h RTX 3060 |
| FP8 (RedHatAI)RedHatAI/Meta-Llama-3.1-8B-FP8 | vLLM | 9.1 GB | 13.5 GB | $0.14/h RTX A4000 |
| Instruct quantized.w8a8 (RedHatAI)RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8 | vLLM | 9.1 GB | 13.5 GB | $0.14/h RTX A4000 |
| unsloth bnb 4bit (unsloth)unsloth/meta-Llama-3.1-8B-unsloth-bnb-4bit | vLLM | 6.0 GB | 10 GB | $0.10/h RTX 3060 |
| Instruct FP8 dynamic (furiosa-ai)furiosa-ai/Meta-Llama-3.1-8B-Instruct-FP8-dynamic | vLLM | 9.1 GB | 13.5 GB | $0.14/h RTX A4000 |
| Instruct int8 (FriendliAI)FriendliAI/Meta-Llama-3.1-8B-Instruct-int8 | vLLM | 9.1 GB | 13.5 GB | $0.14/h RTX A4000 |
| bnb 4bit (unsloth)unsloth/Meta-Llama-3.1-8B-bnb-4bit | vLLM | 5.7 GB | 9.8 GB | $0.10/h RTX 3060 |
| Llama3.1 8B Chinese Chat (shenzhi-wang)shenzhi-wang/Llama3.1-8B-Chinese-Chat | vLLM | 16 GB | 21.3 GB | $0.18/h 2x RTX 3060 |
| Instruct bnb 4bit (unsloth)unsloth/Llama-3.1-8B-Instruct-bnb-4bit | vLLM | 5.7 GB | 9.8 GB | $0.10/h RTX 3060 |
| Instruct pearl (pearl-ai)pearl-ai/Llama-3.1-8B-Instruct-pearl | vLLM | 9.1 GB | 13.5 GB | $0.14/h RTX A4000 |
| Instruct NVFP4 (RedHatAI)RedHatAI/Llama-3.1-8B-Instruct-NVFP4 | vLLM | 6.0 GB | 10.1 GB | $0.10/h RTX 3060 |
| quantized.w8a8 (RedHatAI)RedHatAI/Meta-Llama-3.1-8B-quantized.w8a8 | vLLM | 9.1 GB | 13.5 GB | $0.14/h RTX A4000 |
| llama 3 1 8b base onnx genai int4 kquantlast emb int8 (Arm)Arm/llama-3-1-8b-base-onnx-genai-int4-kquantlast-emb-int8 | ONNX Runtime GenAI | 5.7 GB | 8.6 GB | $0.09/h RTX 3060 |
| llama 3 1 8b instruct onnx genai int4 kquantlast emb int8 (Arm)Arm/llama-3-1-8b-instruct-onnx-genai-int4-kquantlast-emb-int8 | ONNX Runtime GenAI | 5.4 GB | 8.8 GB | $0.09/h RTX 3060 |
| Version | Runtime | Weights | GPU memory | Cheapest machine now |
|---|---|---|---|---|
| Originalmeta-llama/Llama-3.1-70BLicence to accept | vLLM | 141 GB | ~171.6 GB | $1.38/h 8x RTX 3090 |
| Llama 3.3 70B Instructmeta-llama/Llama-3.3-70B-InstructLicence to accept | vLLM | 141 GB | ~171.6 GB | $1.38/h 8x RTX 3090 |
| Instructmeta-llama/Llama-3.1-70B-InstructLicence to accept | vLLM | 141 GB | ~171.6 GB | $1.38/h 8x RTX 3090 |
| llama 3.3 70b instruct awq (casperhansen)casperhansen/llama-3.3-70b-instruct-awq | vLLM | 40 GB | 49.4 GB | $0.78/h 4x RTX 4060 Ti |
| Instruct FP8 (RedHatAI)RedHatAI/Meta-Llama-3.1-70B-Instruct-FP8 | vLLM | 73 GB | 85.9 GB | $0.89/h 2x Nvidia A40 |
| Llama 3.3 70B Instruct FP8 dynamic (RedHatAI)RedHatAI/Llama-3.3-70B-Instruct-FP8-dynamic | vLLM | 73 GB | 85.9 GB | $0.89/h 2x Nvidia A40 |
| Meta Llama 3.3 70B Instruct AWQ INT4 (ibnzterrell)ibnzterrell/Meta-Llama-3.3-70B-Instruct-AWQ-INT4 | vLLM | 40 GB | 49.4 GB | $0.78/h 4x RTX 4060 Ti |
| Llama 3.3 70B Instruct bnb 4bit (unsloth)unsloth/Llama-3.3-70B-Instruct-bnb-4bit | vLLM | 40 GB | 49.1 GB | $0.78/h 4x RTX 4060 Ti |
| Llama 3.3 70B Instruct quantized.w4a16 (RedHatAI)RedHatAI/Llama-3.3-70B-Instruct-quantized.w4a16 | vLLM | 40 GB | 49.1 GB | $0.78/h 4x RTX 4060 Ti |
| Instruct quantized.w4a16 (RedHatAI)RedHatAI/Meta-Llama-3.1-70B-Instruct-quantized.w4a16 | vLLM | 40 GB | 49.4 GB | $0.78/h 4x RTX 4060 Ti |
| Version | Runtime | Weights | GPU memory | Cheapest machine now |
|---|---|---|---|---|
| Originalmeta-llama/Llama-3.1-405B-FP8Licence to accept | vLLM | 487 GB | ~586.9 GB | $11.77/h 8x RTX PRO 6000 |
| Version | Runtime | Weights | GPU memory | Cheapest machine now |
|---|---|---|---|---|
| Originalmeta-llama/Llama-3.1-405BLicence to accept | vLLM | 812 GB | ~976.3 GB | $44.03/h 8x Nvidia H200 |
| Instructmeta-llama/Llama-3.1-405B-InstructLicence to accept | vLLM | 812 GB | 908.8 GB | $44.03/h 8x Nvidia H200 |
| Instruct FP8meta-llama/Llama-3.1-405B-Instruct-FP8Licence to accept | vLLM | 487 GB | 548.3 GB | $11.77/h 8x RTX PRO 6000 |
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").