Basic information
Model repository: deepseek-ai/DeepSeek-V3
Organization: deepseek-ai
Tasks: text-generation
Architecture and runtime
Model type: deepseek_v3
Architecture: Not provided
Libraries: safetensors, pytorch
Base models: Not provided
Current revision: master
License and update history
License: Not provided
Created: 2025-09-18
Last updated: 2025-09-18
Hardware requirements
参考配置（非官方最低要求）：官方模型卡列出总参数 671B，每个 token 激活参数约 37B。仓库提供的 FP8 权重（含 MTP）约 685 GB，建议使用多卡分布式推理，FP8 合计显存按约 800 GB 起步估算，并为 KV cache 预留余量。BF16 权重约 1.37 TB。以上均为估算，非官方最低配置。 Inference reference: the official card lists 671B total and about 37B active parameters per token; the repository provides FP8 weights (about 685 GB including MTP weights). Use distributed multi-GPU inference with at least about 800 GB aggregate VRAM for FP8 plus KV-cache headroom. BF16 weights are roughly 1.37 TB. These are weight-size estimates, not official minimum requirements.
Download links
Basic usage example
启动 vLLM 多卡服务 / Start multi-GPU vLLM server
# Example for a multi-GPU host; size parallelism to the available hardware.
pip install -U "vllm>=0.23.0" openai
vllm serve deepseek-ai/DeepSeek-V3 \
--tensor-parallel-size 8 \
--pipeline-parallel-size 2 \
--max-model-len 32768 \
--trust-remote-code调用 OpenAI 兼容接口 / Call the OpenAI-compatible API
from openai import OpenAI
client = OpenAI(api_key="EMPTY", base_url="http://localhost:8000/v1")
response = client.chat.completions.create(
model="deepseek-ai/DeepSeek-V3",
messages=[{"role": "user", "content": "请简要解释混合专家模型（MoE）的工作方式。"}],
max_tokens=512,
)
print(response.choices[0].message.content)