Llama 3 models
19 Llama 3 models from Meta on Hugging Face, from 8.0B to 70.6B parameters, published by Meta in BF16, with a 8K-token context. The smallest official model, Meta-Llama-3-8B, needs about 17.9 GB of VRAM at its published precision; the cheapest live fit is RTX A5000 at $0.176/hr.
Part of the Llama series · Previous generation: Llama 2 · Next generation: Llama 3.1
Pick a size
One row per official Llama 3 size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.
| Model | Parameters | Native VRAM | FP8 VRAM | INT4 VRAM | Live GPU fit (native) | Est. $/hr | KV cache at 32K |
|---|---|---|---|---|---|---|---|
| Meta-Llama-3-8B-Instruct | 8.0B | 17.9 GB | 9.0 GB | 4.5 GB | RTX A5000 | $0.176/hr | 4.0 GB |
| Meta-Llama-3-70B | 70.6B | 158 GB | 78.8 GB | 39.4 GB | AMD MI300X | $2.63/hr | 10.0 GB |
VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.
Official models (4)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| Meta-Llama-3-8B | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| Meta-Llama-3-8B-Instruct | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| Meta-Llama-3-70B | 70.6B | BF16 | 158 GB | AMD MI300X | $2.63/hr |
| Meta-Llama-3-70B-Instruct | 70.6B | BF16 | 158 GB | AMD MI300X | $2.63/hr |
Fine-tunes and community models (15)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| Bio-Medical-Llama-3-8B | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| Llama-3-8b-Ita | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| Llama3-Chinese-8B-Instruct | 8.0B | F16 | 17.9 GB | RTX A5000 | $0.176/hr |
| Llama-3-8B-Instruct-Gradient-1048k | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| finance-Llama3-8B | 8.0B | F32 | 35.9 GB | RTX A6000 | $0.363/hr |
| Tess-2.0-Llama-3-8B | 8.0B | F16 | 17.9 GB | RTX A5000 | $0.176/hr |
| NeuralDaredevil-8B-abliterated | 8.0B | F16 | 17.9 GB | RTX A5000 | $0.176/hr |
| openchat-3.6-8b-20240522 | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| Llama-3-Patronus-Lynx-8B-Instruct-v1.1 | 8.0B | F16 | 17.9 GB | RTX A5000 | $0.176/hr |
| Llama-3-8B-ProLong-64k-Base | 8.0B | F32 | 35.9 GB | RTX A6000 | $0.363/hr |
| Llama-3-8B-UltraMedical | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| llama-3-8b | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| llama-3-8b-Instruct | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| Llama3-ChatQA-1.5-8B | 8.0B | F16 | 17.9 GB | RTX A5000 | $0.176/hr |
| Llama-3-Taiwan-8B-Instruct | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.
Run and fine-tune Llama 3
- How to run Llama 3: engine support, chat template and context flags, with a serving command sized per model.
- How to fine-tune Llama 3: LoRA and QLoRA memory per model, and the license limits on derivatives.
More from Meta
- Llama 3.2 (15 models)
- Llama 3.1 (40 models)
- Llama 2 (8 models)
- Llama 3.3 (4 models)
- Code Llama (3 models)
- Llama Guard (4 models)
- Llama 4 (2 models)
- Llama 1 (1 models)