DeepSeek-V3: GPU memory and hosting cost
Open weights · 684.53 billion parameters · Mixture of experts · 163,840-token context · licence mit
No cloud sells this model per token: to use it you run it yourself, on GPU instances you rent.
Run it yourself
It needs 855.7 GB of GPU memory at fp8, the precision it ships in.
| Cloud | Cheapest instance that fits | GPUs | An hour | A month, around the clock |
|---|---|---|---|---|
| Crusoe | mi300x.8x | 8 × MI300X 192 GB | $27.60 | $20,148 |
| Nebius | gpu-h200-sxm.8gpu-128vcpu-1600gb | 8 × H200 141 GB | $43.20 | $31,536 |
| Together AI | hgx-h200.8gpu | 8 × H200 141 GB | $47.92 | $34,982 |
| OCI | BM.GPU.MI300X.8 | 8 × MI300X 192 GB | $48.00 | $35,040 |
| AWS | 2 × p4de.24xlarge, joined by EFA | 8 × A100 80 GB in each | $54.89 | $40,073 |
| Azure | 2 × Standard_ND96amsr_A100_v4, joined by InfiniBand | 8 × A100 80 GB in each | $65.54 | $47,844 |
| Google Cloud | a3-ultragpu-8g | 8 × H200 141 GB | $84.81 | $61,909 |
| Alibaba Cloud | 2 × ecs.gn8v-8x.16xlarge, joined by eRDMA | 8 × GPU H 96 GB in each | $116.98 | $85,397 |
Several servers: the model is sized across identical servers that hold it together, on instance types whose vendor documents a network between servers; the price is all of them. It is shown where no one server holds the model, or where it costs less than the one that does.
More from Unsloth
- Llama-3.2-1B-Instruct (unsloth)
- Llama-3.2-3B-Instruct (unsloth)
- Llama-3.1-8B-Instruct (unsloth)
- Meta-Llama-3.1-8B-Instruct (unsloth)
- Qwen2.5-Coder-7B-Instruct (unsloth)
- Qwen2.5-7B-Instruct (unsloth)
Similar size, open weights
List prices collected 4 November 2026, refreshed every three hours. No negotiated rates or taxes.