What runs this LLM?
Pick the model you want to serve. We size it from the released checkpoint and list the cheapest configurations you can rent today that hold it — with the on-demand and serverless rate for each.
1 · Choose a model
Released at 4-bit only — never sized above the checkpoint it shipped as.
Context-length sizing isn’t published for this architecture.
2 · What it needs
DeepSeek V4 Flash at 4-bit
192 GB of VRAM
at short context — measured from the released checkpoint, not estimated from parameter count.
- Model weights
- 160 GB
- KV cachefolded into overhead
- —
- Runtime overhead20% of weights
- 32 GB
About this model
- Parameters
- 284B
- Released at
- 4-bit
- Measured weights
- 160 GB
3 · Where to run it
Bookable configurations
Cheapest first, checked against 98 GPU types with live bookable capacity.
| Configuration | Total VRAM | On-demand | Serverless | GPU details |
|---|---|---|---|---|
| 8× NVIDIA RTX 3090 | 192 GB 0.5 GB free | $0.72/hr $0.09/GPU-hr  Vast.ai | — | See all prices |
| 12× NVIDIA RTX A4000 | 192 GB 0.5 GB free | $0.96/hr $0.08/GPU-hr Vast.ai | — | See all prices |
| 4× NVIDIA Quadro RTX 8000 | 192 GB 0.5 GB free | $1.00/hr $0.25/GPU-hr Vast.ai | — | See all prices |
| 4× NVIDIA RTX A6000 | 192 GB 0.5 GB free | $1.68/hr $0.42/GPU-hr  Lium | — | See all prices |
| 10× NVIDIA RTX 5060 TiTight fit | 160 GB 32 GB short of the margin | $1.70/hr $0.17/GPU-hr Vast.ai | — | See all prices |
| 1× AMD Instinct MI300X | 192 GB 0.5 GB free | $1.71/hr  TensorWave | — | See all prices |
| 4× NVIDIA L40S | 192 GB 0.5 GB free | $1.88/hr $0.47/GPU-hr Vast.ai | $6.03/hr $1.51/GPU-hr  Verda | See all prices |
| 2× Intel Gaudi 2 | 192 GB 0.5 GB free | $2.12/hr $1.06/GPU-hr  Cyfuture AI | — | See all prices |
Showing the 8 cheapest. Model weights plus 20% for runtime overhead, from measured checkpoint sizes; a stated context length adds its KV cache on top. Each price is the cheapest bookable rate for that exact instance size, for the whole configuration — the cheaper of the two is highlighted and orders the table. A ×N figure is total VRAM across N cards and makes no claim about interconnect throughput.
How this is worked out
Weight sizes are measured from the actual checkpoint files on Hugging Face — never derived from parameter count, which gets quantized releases wrong. On top of the weights we assume 20% for activations, CUDA context and allocator slack. Stating a context length adds its KV cache explicitly, computed only for architectures whose cache shape the config states unambiguously.
A configuration that holds the weights and cache but not that full margin is marked as a tight fit rather than hidden — it will load, and may well run at batch size one, but expect trouble under load.
Prices are the cheapest bookable rate for that exact instance size, refreshed with the rest of the catalog.