Models · Meta · live prices
Llama 3.1 8B: GPUs needed and cost to self-host
8B parameters. In BF16 that is about 19 GB with room for the KV cache.
Memory, BF16
19 GB
weights + ~20%
Cheapest per 1M output tokens
$0.05
1× A100 · on-demand · 60% utilization
Hosted API, sample list price
$0.18
per 1M output tokens
Cost per 1M output tokens, BF16
Prices observed 12m ago. Cheapest on-demand offer per GPU, whole instances, 60% average utilization.
- A10080 GB 1× on Vast.ai · on-demand · $0.34/hr · ~3,059 tok/s $0.05per 1M out
- 509032 GB 1× on Vast.ai · on-demand · $0.46/hr · ~2,688 tok/s $0.08per 1M out
- H200141 GB 1× on Vast.ai · on-demand · $1.32/hr · ~7,200 tok/s $0.08per 1M out
- 409024 GB 1× on Vast.ai · on-demand · $0.31/hr · ~1,512 tok/s $0.10per 1M out
- RTX PRO 600096 GB 1× on Vast.ai · on-demand · $1.00/hr · ~2,688 tok/s $0.17per 1M out
- H10080 GB 1× on RunPod · community · $1.99/hr · ~5,025 tok/s $0.18per 1M out
- B300288 GB 1× on RunPod · community · $6.94/hr · ~12,000 tok/s $0.27per 1M out
The tick marks the sample hosted-API price. Estimates: tokens/s from memory bandwidth with batched serving, not a benchmark. Idle hours still bill, so lower utilization means pricier tokens.
GPUs per replica by precision
One node, split 1, 2, 4 or 8 ways; 90% of VRAM usable.
| GPU | BF1619 GB | FP810 GB | INT45 GB |
|---|---|---|---|
| H10080 GB | 1× | 1× | 1× |
| H200141 GB | 1× | 1× | 1× |
| B300288 GB | 1× | 1× | 1× |
| A10080 GB | 1× | 1× | 1× |
| RTX PRO 600096 GB | 1× | 1× | 1× |
| 509032 GB | 1× | 1× | 1× |
| 409024 GB | 1× | 1× | 1× |