DeepSeek V4 Flash GPU requirements: 167 GB at FP4+FP8, 608 GB at BF16
DeepSeek V4 Flash needs 608 GB of GPU memory for its weights at BF16, 304 GB at FP8 and 167 GB at FP4+FP8. The cheapest node that runs it here is 2× RTX PRO 6000 Blackwell at FP4+FP8 (native), for $0.38 an hour on spot.
Weights plus a 32K 16-bit KV cache, with 10% of the node's memory kept free. Billed per minute.
- Parameters
- 284B13B active per token
- Weights, FP4+FP8
- 167 GBBF16: 608 GB
- KV cache
- CompressedNot estimated, see below
- Context window
- 1MNative
How much VRAM DeepSeek V4 Flash needs, format by format.
Weights are parameters × bytes per parameter, except the native FP4+FP8 row, which is the size of the published checkpoint. The KV cache of this architecture is not estimated (see below), so only the weights are shown.
| Format | Weights | Cheapest node, 32K |
|---|---|---|
| BF16Published weights, fine-tuning | 608 GB | 8× RTX PRO 6000 Blackwell · $1.52/h |
| FP8Hopper, Ada, Blackwell, MI300 | 304 GB | 8× L40 · $0.72/h |
| FP4+FP8 (native)AWQ, GPTQ, GGUF Q4 · the published checkpoint | 167 GB | 2× RTX PRO 6000 Blackwell · $0.38/h |
The serving engine adds its own buffers, and an 8-bit KV cache halves the context columns. We keep 10% of a node's memory free when we say a model fits. Speed is a separate question: it follows memory bandwidth and, for a mixture of experts, the active parameters.
GPUs that run DeepSeek V4 Flash, cheapest first.
For each format, the smallest node of each GPU model that holds the weights and a 32K context, sorted by spot price. One row per GPU model; larger nodes of the same GPU work too.
| Format | GPU | GPU memory | Spot | On-demand | 730 h spot | Action |
|---|---|---|---|---|---|---|
| FP4+FP8 (native) | 2× RTX PRO 6000 BlackwellBlackwell · 167 GB needed | 192 GB | $0.38 | $0.78 | $277.40 | Deploy |
| FP4+FP8 (native) | 4× RTX PRO 5000 BlackwellBlackwell · 167 GB needed | 192 GB | $0.60 | $1.16 | $438.00 | Deploy |
| FP4+FP8 (native) | 2× B200 SXMBlackwell · 167 GB needed | 360 GB | $3.38 | $4.98 | $2,467.40 | Deploy |
| FP8 | 8× L40Ada Lovelace · 304 GB needed | 384 GB | $0.72 | $1.52 | $525.60 | Deploy |
| FP8 | 4× RTX PRO 6000 BlackwellBlackwell · 304 GB needed | 384 GB | $0.76 | $1.56 | $554.80 | Deploy |
| FP8 | 8× L40SAda Lovelace · 304 GB needed | 384 GB | $1.20 | $2.00 | $876.00 | Deploy |
| FP8 | 4× H100 NVLHopper · 304 GB needed | 376 GB | $1.96 | $3.16 | $1,430.80 | Deploy |
| BF16 | 8× RTX PRO 6000 BlackwellBlackwell · 608 GB needed | 768 GB | $1.52 | $3.12 | $1,109.60 | Deploy |
| BF16 | 8× H100 NVLHopper · 608 GB needed | 752 GB | $3.92 | $6.32 | $2,861.60 | Deploy |
| BF16 | 4× MI300XCDNA 3 · 608 GB needed | 768 GB | $3.96 | $5.80 | $2,890.80 | Deploy |
| BF16 | 4× MI325XCDNA 3 · 608 GB needed | 1024 GB | $4.36 | $6.36 | $3,182.80 | Deploy |
Spot capacity can be reclaimed with a 2-minute notice; for a live endpoint, use on-demand. How spot works · serving a language model.
What a long context costs in memory.
Compressed attention (per-layer compression ratios of 4 and 128 plus a 128-token window); its cache is a small fraction of a standard model's and we publish no estimate we cannot check.
The cache is a small fraction of a standard model's at the same length, but its size depends on details we cannot check from the published configuration, so we do not print a number. The node recommendations on this page are sized on the weights plus the 10% margin.
Serving several users
KV cacheEvery concurrent request keeps its own cache, so memory for context grows with users × context length. A server batching 8 requests of 32K needs eight times the single-sequence figure.
Paged attention (vLLM, SGLang) allocates the cache in blocks as requests grow, so a node holds more short requests than the worst case suggests.
DeepSeek V4 Flash, as published.
- Developer
- DeepSeekFast DeepSeek V4 mixture of experts with a 1M-token context, the 0731 update
- Parameters
- 284B total, 13B active per token304B in the published checkpoint (embedding tables, prediction or vision modules included); memory figures use this count
- Architecture
- Mixture of experts, 43 layersCompressed attention (per-layer compression ratios of 4 and 128 plus a 128-token window); its cache is a small fraction of a standard model's and we publish no estimate we cannot check
- Context window
- 1,048,576 tokensAs set in the published configuration
- Published format
- FP4+FP8, 167 GB checkpoint
- Licence
- MIT
- Released
- 31 July 2026
- Source
- deepseek-ai/DeepSeek-V4-Flash-0731config.json read 26 October 2026
Serve it with vLLM
template: vLLMOn 2× RTX PRO 6000 Blackwell, from the instance terminal:
$ vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \ --tensor-parallel-size 2 \ --max-model-len 32768
Very recent architectures need a recent vLLM; the model card gives the minimum version and any extra flags.
DeepSeek V4 Flash on a GPU — the questions that come up
How much VRAM does DeepSeek V4 Flash need?
For the weights alone: 608 GB at BF16, 304 GB at FP8 and 167 GB at FP4+FP8. Its attention scheme keeps a much smaller cache than a standard model and we do not publish an estimate we cannot check; keep a margin for the cache on top of the weights.
Can DeepSeek V4 Flash run on a single GPU?
One RTX 4090 (24 GB): no. One RTX 5090 (32 GB): no. One H100 SXM (80 GB): no. One H200 SXM (141 GB): no. One B200 SXM (180 GB): no. Those answers assume a 32K context and the widest format that still fits.
What is the cheapest way to run DeepSeek V4 Flash?
Here, 2× RTX PRO 6000 Blackwell at FP4+FP8 (native): $0.38 per hour on spot and $0.78 on-demand, billed per minute — $277.40 for a 730-hour month on spot. That node has 192 GB of GPU memory for the 167 GB this setup needs.
Does DeepSeek V4 Flash need less memory because only 13B parameters are active?
No. A mixture-of-experts model routes each token through a few experts, but every expert has to be resident in GPU memory, so the memory bill follows the 284B total. What the 13B active parameters change is the work per token: generation reads far less of the weights each step than a dense model of the same size would, which is why it is fast for its size.
Which format should I run it in?
DeepSeek V4 Flash is published in FP4+FP8, a 167 GB checkpoint — the format its developers serve it in, so it is the safe default. BF16 keeps the published weights untouched and is the format for fine-tuning. FP8 halves the memory and runs in hardware on Hopper, Ada, Blackwell and MI300-class GPUs. 4-bit quantisation (AWQ, GPTQ, GGUF) quarters it, at a quality cost that is larger on small models than on big ones; check it against your own evaluation before relying on it.
How do I deploy it?
Pick the node in the table above and launch it with the vLLM template; the model downloads from Hugging Face (deepseek-ai/DeepSeek-V4-Flash-0731) on first start. Spread the weights over the node with --tensor-parallel-size equal to the GPU count, and set --max-model-len to the context you need. New architectures need a recent vLLM: check the model card for the minimum version.
Sizing another model? GPU requirements for 64 open models · serving a language model · fine-tuning.
DeepSeek V4 Flash on 2× RTX PRO 6000 Blackwell for $0.38 an hour.
Launch the node with the vLLM template, point it at deepseek-ai/DeepSeek-V4-Flash-0731, stop it when you are done. Prepaid, billed by the minute.
Billed per minute from the moment the instance is reachable. Minimum credit $30, no subscription.