What GPU do I need to run deepseek-ai/DeepSeek-V4-Flash-0731?
A 284B-parameter (13B active) reasoning model with a 1M-token context. 304.2B parameters, published in FP4 + FP8 Mixed. View on Hugging Face
What DeepSeek-V4-Flash-0731 is
DeepSeek-V4-Flash-0731 is DeepSeek's official release of DeepSeek-V4-Flash, a 284B-parameter mixture-of-experts reasoning model with 13B active parameters, the smaller sibling of DeepSeek-V4-Pro with a DSpark speculative-decoding module attached. Its own model card states the same 1,048,576-token (1M) context window, hybrid attention architecture, three reasoning-effort levels (low, high, max), and mixed FP4 (MoE experts) + FP8 (other parameters) published precision as V4-Pro.
License note: fully permissive, no gating and no usage restrictions. Facts in this section are sourced from DeepSeek-V4-Flash-0731's Hugging Face model card, not benchmarked by Aquanode.
What it's used for
- Multi-step reasoning and agentic coding
- Long-context analysis (up to 1M tokens)
- Lower-latency deployment than V4-Pro
Benchmarks (published by DeepSeek)
Published by DeepSeek, not measured by Aquanode.
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision times a flat 1.2 overhead for activations and fragmentation. The KV-cache grows with context and is not in that factor; it is listed per model below where the architecture is published. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| FP4 + FP8 Mixed (native) | 155.4 GB | 564.0 GB | AMD MI300X | 3 | $7.89/hr |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
The SGLang DeepSeek-V4 cookbook linked from DeepSeek's model card lists DeepSeek-V4-Flash-0731 as verified on 4×H200; NVIDIA's own H200 spec lists 141GB HBM3e per GPU, so 4×141GB ≈ 564GB total.
Cheapest way to run DeepSeek-V4-Flash-0731 at its published (FP4 + FP8 Mixed) precision: 3× AMD MI300X, at $2.63/hr per GPU ($7.89/hr total).
DeepSeek-V4-Flash-0731: KV cache by context length
The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.
KV-cache size for DeepSeek-V4-Flash-0731 is not published for this architecture, so no figure is shown.
How to run DeepSeek-V4-Flash-0731
Run DeepSeek-V4-Flash-0731 with vLLM
From DeepSeek-V4-Flash-0731's own model card: an example serving it with vLLM and DSpark speculative decoding on a single 4×GB300 node.
vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \
--trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
--data-parallel-size 4 --enable-expert-parallel \
--moe-backend deep_gemm_mega_moe \
--attention-config '{"use_fp4_indexer_cache": true}' \
--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731/raw/main/README.md
Run DeepSeek-V4-Flash-0731 with SGLang
From DeepSeek-V4-Flash-0731's own model card, same single 4×GB300 node example as the vLLM command above.
sglang serve \
--trust-remote-code \
--model-path deepseek-ai/DeepSeek-V4-Flash-0731 \
--tp 4 \
--moe-runner-backend flashinfer_mxfp4 \
--speculative-algorithm DSPARK \
--mem-fraction-static 0.90 \
--chunked-prefill-size 4096Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731/raw/main/README.md
Deploy DeepSeek-V4-Flash-0731 on Aquanode
Aquanode has no one-click deploy template for DeepSeek-V4-Flash-0731; you install the inference engine yourself with the commands below. Aquanode sells GPU pods billed per second, not a hosted inference API.
- Launch a bare GPU pod sized to the requirement above (3× AMD MI300X or larger).
- Open a terminal on the pod, or save one of the commands above as a startup script so it runs automatically the first time the pod boots.
- Run the command and connect to the resulting endpoint.
Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More DeepSeek V4 models
- DeepSeek-V4.1-Flash-1017-obliterated (754.6B, BF16)
- DeepSeek-V4-Pro-0813 (1650.5B, FP4 + FP8 Mixed)
Alternatives at this size
Other models for text-generation within about a third of DeepSeek-V4-Flash-0731's 304.2B parameters, from other model lines.
- MiniMax-M2.7 (228.7B, F8_E4M3)
- MiMo-V2.5 (310.8B, F8_E4M3)
- Qwen3-235B-A22B (235.1B, BF16)
- Ornith-1.5-397B (403.4B, BF16)
- Llama-3.1-405B (405.9B, BF16)
More on DeepSeek-V4-Flash-0731
Best chat and assistants models: how DeepSeek-V4-Flash-0731 ranks against the rest.
Related reading: AMD MI300X pricing and specs, How much VRAM you need for LLMs, Serving LLMs with vLLM, vLLM vs TensorRT-LLM vs SGLang, and Best GPU for LLM inference.
Your own GPU cloud accounts
One console for the GPU clouds you already rent from. Your keys stay on your machine.
$20 a month, flat