Models · Mistral · live prices
Mixtral 8x22B: GPUs needed and cost to self-host
141B parameters, 39B active per token (mixture of experts: every expert sits in memory, but only the active ones are read per token, so it decodes like a 39B model). In FP8 that is about 169 GB with room for the KV cache.
Memory, FP8
169 GB
weights + ~20%
Cheapest per 1M output tokens
$0.12
4× A100 · on-demand · 60% utilization
Hosted API, sample list price
$1.20
per 1M output tokens
Cost per 1M output tokens, FP8
Prices observed 11m ago. Cheapest on-demand offer per GPU, whole instances, 60% average utilization.
- A10080 GB 4× on Vast.ai · on-demand · $1.34/hr · ~5,019 tok/s $0.12per 1M out
- H200141 GB 2× on Vast.ai · on-demand · $2.64/hr · ~5,908 tok/s $0.21per 1M out
- 509032 GB 8× on Vast.ai · on-demand · $3.66/hr · ~5,293 tok/s $0.32per 1M out
- 409024 GB 8× on Vast.ai · on-demand · $2.49/hr · ~2,977 tok/s $0.39per 1M out
- H10080 GB 4× on RunPod · community · $7.96/hr · ~8,246 tok/s $0.45per 1M out
- B300288 GB 1× on RunPod · community · $6.94/hr · ~4,923 tok/s $0.65per 1M out
- RTX PRO 600096 GB 2× on Vast.ai · on-demand · $2.00/hr · ~1,323 tok/s $0.70per 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 | BF16338 GB | FP8169 GB | INT485 GB |
|---|---|---|---|
| H10080 GB | 8× | 4× | 2× |
| H200141 GB | 4× | 2× | 1× |
| B300288 GB | 2× | 1× | 1× |
| A10080 GB | 8× | 4× | 2× |
| RTX PRO 600096 GB | 4× | 2× | 1× |
| 509032 GB | — | 8× | 4× |
| 409024 GB | — | 8× | 4× |