Run DeepSeek V4 on cloud GPUs

The DeepSeek V4 model that fits an ordinary multi-GPU rental is V4-Flash. Its weights are 166.89 GB, so a 4x H200 VM holds them with about 352 GB left over. V4-Pro is 892.74 GB and needs all 8 GPUs of an H200 VM, and V4.1-Flash at 510.30 GB sits between them. All three are MIT licensed, and Ollama offers none of them as a local download.

Three V4 checkpoints: 167, 510 and 893 GB

The V4 checkpoints are mixed precision, with the experts stored in 4-bit FP4 and the rest in FP8, so they are small for their parameter counts. The cards list V4-Flash at 284 billion parameters with 13 billion active, V4-Pro at 1.6 trillion with 49 billion active, and V4.1-Flash at 552 billion backbone parameters, all with a 1M-token context. The context is cheap to hold. vLLM's launch post puts the BF16 KV cache at 9.62 GiB per 1M-token sequence, against 83.9 GiB for a 61-layer V3.2-style stack, and the KV cache explainer shows how such figures are worked out.

The left-over figures are calculated as 92% of the VM's memory minus the checkpoint. They leave out activations and CUDA graphs, and the engine's startup log has the real KV cache size. Sizes are the summed safetensors files and every card, recipe and cookbook in this article was checked on 1 October 2026.

V4-Flash on a 4x H200 VM

My starting point for V4 is V4-Flash-0731 on four H200s. SGLang's cookbook serves V4-Flash on a single H200 node at TP 4, and vLLM's recipe lists H200 among its verified hardware. An H200 SXM board joins GPUs through NVSwitch and an NVL bridge joins 2 or 4 cards, so check the Interconnect column below and run nvidia-smi topo -m on the VM.

Here are the essentials of the cookbook's H200 setup at TP 4: SGLang's Hopper image for V4 and the MoE runner that handles the FP4 experts on Hopper, wrapped in Docker. Check the VM's driver with nvidia-smi before pulling it. The port is published on the VM's loopback address and the --host 0.0.0.0 stays inside the container:

sudo docker run -d --name sglang --gpus all --shm-size 32g --ipc=host \
  -p 127.0.0.1:30000:30000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  lmsysorg/sglang:deepseek-v4-hopper \
  sglang serve \
  --trust-remote-code \
  --model-path deepseek-ai/DeepSeek-V4-Flash-0731 \
  --tp 4 \
  --moe-runner-backend flashinfer_mxfp4 \
  --host 0.0.0.0 \
  --port 30000

The cookbook's configurator on that page generates the tuned profiles, such as DSpark speculative decoding.

For vLLM, the recipe's command builder prints the exact flags for your hardware. The command below starts from the vLLM launch post's V4-Flash command, written for 4x B200 or B300. It drops the FP4 indexer flag, which the recipe uses only on Blackwell, and uses the image and model name the recipe gives for the 0731 checkpoint:

sudo docker run -d --name vllm --gpus all --ipc=host \
  -p 127.0.0.1:8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:v0.28.0 deepseek-ai/DeepSeek-V4-Flash-0731 \
  --trust-remote-code \
  --kv-cache-dtype fp8 \
  --block-size 256 \
  --enable-expert-parallel \
  --data-parallel-size 4 \
  --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE", "custom_ops":["all"]}' \
  --tokenizer-mode deepseek_v4 \
  --tool-call-parser deepseek_v4 \
  --enable-auto-tool-choice \
  --reasoning-parser deepseek_v4

Reach either server through an SSH tunnel, as in the vLLM guide and the SGLang guide. The recipe asks for vLLM 0.25.0 or newer for the 0731 checkpoint, and Think Max needs a context of at least 393,216 tokens, so do not cap --max-model-len below that if you use it.

V4.1-Flash and V4-Pro need all 8 GPUs

V4.1-Flash is the newest of the three. vLLM added it in version 0.30.0, released on 22 September 2026, and its recipe sizes the model at 614 GB, a fit for 8x H200. SGLang's V4.1 cookbook has launch recipes for H200, B200 and B300: copy the image and command its configurator generates. Four H200s hold the weights but leave about 9 GB, so use eight.

V4-Pro is where the memory math bites. vLLM's recipe sizes it at 960 GB, and SGLang's cookbook lists 8x H200 at TP 8 for the 0813 checkpoint. Eight H100s hold 640 GB, which is not enough, and SGLang's H100 recipe uses 2 nodes. DeployGPU has no multi-node VMs, so that route goes through the cluster quote form. The frontier models page sets V4-Pro beside Kimi K3 and GLM-5.3.

DeepSeek R1 and its distills

R1 is the older reasoning model, released in January 2025. Its model card lists 671 billion parameters, 37 billion active and a 128K context under the MIT license. It shares its architecture with DeepSeek-V3, so vLLM's R1 recipe covers both and names 8x H200 for the FP8 weights. That checkpoint is 688.59 GB, which leaves about 349 GB on 8x H200 (calculated). The card also publishes six distilled dense models from 1.5B to 70B. The 32B Qwen distill is 65.53 GB in BF16, which fits one 96 GB RTX PRO 6000 with about 23 GB left (calculated), or two 48 GB cards. The card's own vLLM command splits it over 2 GPUs. In Docker, with the port on loopback, it reads:

sudo docker run -d --name vllm --gpus all --ipc=host \
  -p 127.0.0.1:8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:latest \
  --model deepseek-ai/DeepSeek-R1-Distill-Qwen-32B \
  --tensor-parallel-size 2 \
  --max-model-len 32768 \
  --enforce-eager

For chat, Ollama lists deepseek-r1:32b at 20 GB, deepseek-r1:70b at 43 GB and deepseek-r1:671b at 404 GB, checked 1 October 2026. The 32B tag runs on a 48 GB card:

Ollama and llama.cpp for V4

Ollama has a deepseek-v4-flash library page with no models pushed, while deepseek-v4-pro and deepseek-v4.1-flash are cloud tags only. So on a rented GPU the engines are vLLM and SGLang. llama.cpp registers a converter for the V4 architecture but not for V4.1, and DeepSeek's Hugging Face organization lists no GGUF files. The llama.cpp server guide covers the engine for models that ship as GGUF.

MIT licenses and a 167 GB download

V4-Flash, V4-Pro, V4.1-Flash and R1 are all MIT licensed, which allows commercial use and fine-tuning. The R1 distills inherit their base models' terms, which are Apache 2.0 for the Qwen distills and the Llama 3.1 and 3.3 licenses for the Llama ones. Stopping a DeployGPU instance deletes its disk, so every new session downloads the weights again: 167 GB for V4-Flash and 893 GB for V4-Pro. The stop guide shows what to copy off first.