D
DeepSeek V3.2
DeepSeek · DeepSeek V3.2 (Reasoning)
轻量模型，代码能力强
Input
¥2.0
/M Tokens
Output
¥3.0
/M Tokens
Context
128K
Tokens
Capabilities
对话代码
Best For
简单问答
查看官方定价页面
M
在 ModelSage 接入此模型Quick Cost Estimate
¥19.0/月
10,000 次调用 · 输入 500 tokens · 输出 300 tokens
自建可行性
vLLM 官方 · 2026-10-14
DeepSeek V3.2 私有化部署，NVFP4 量化下需显存约 403GB， 至少 3 张 H200，聚合吞吐约 14,033 tok/s。
NVFP4403GB
3×H200 / 6×H100 SXM
REDHAT-NVFP4430GB
4×H200 / 6×H100 SXM
FP8805GB
6×H200 / 2×DGX Station (GB300)
vLLM 官方验证过的硬件
H100 SXMH200
同样的量走 API：日均 2 亿 token 约 ¥13,500/月。自建是否更便宜取决于卡价与实际利用率。
按实际卡价测算回本周期显存需求与硬件兼容性取自 vLLM 官方 recipes；卡数按「装得下」估算， 未计入张量并行度、互联拓扑与跨节点开销，实际只多不少。吞吐为公式估算值，非实测。
Price Comparison同价位的其他厂商，点进去看详情
DeepSeek V3.2CURRENT
¥2.0 / ¥3.0INPUT
OUTPUT
Qwen3.6 Plus
¥2.0 / ¥12.0INPUT
OUTPUT
Qwen3.7 Plus
¥2.0 / ¥8.0INPUT
OUTPUT
Qwen3 235B A22B 2507
¥2.0 / ¥20.0INPUT
OUTPUT
Qwen3 235B 2507
¥2.0 / ¥8.0INPUT
OUTPUT
DeepSeek Models
DeepSeek V4.1 Flash (max)
¥2.0 / ¥8.0DeepSeek V4 Pro 0813 (max)
¥9.0 / ¥27.0DeepSeek V4 Flash Vision (max)
¥2.9 / ¥8.8DeepSeek V4 Flash 0731 (max)
¥2.0 / ¥8.0DeepSeek V4 Pro (max)
¥9.0 / ¥27.0DeepSeek V4 Pro (high)
¥9.0 / ¥27.0DeepSeek V4.1 Flash (non-reasoning)
¥2.0 / ¥8.0DeepSeek V4 Flash (high)
¥2.0 / ¥8.0DeepSeek V4 Flash (max)
¥2.0 / ¥8.0DeepSeek V4 Pro (non-reasoning)
¥9.0 / ¥27.0DeepSeek V4 Pro 0813 (non-reasoning)
¥9.0 / ¥27.0DeepSeek V4 Flash (non-reasoning)
¥2.0 / ¥8.0DeepSeek V3.2 Exp
¥2.0 / ¥3.0DeepSeek V3.2 (non-reasoning)
¥2.0 / ¥3.0DeepSeek V3.1 Terminus
¥4.0 / ¥12.0DeepSeek V3.1 Terminus (non-reasoning)
¥4.0 / ¥12.0DeepSeek V3.2 Exp (non-reasoning)
¥2.0 / ¥3.0DeepSeek V3.1 (non-reasoning)
¥4.0 / ¥12.0DeepSeek V3.1
¥4.0 / ¥12.0DeepSeek R1 0528
¥4.0 / ¥16.0DeepSeek R1 (Jan)
¥4.0 / ¥16.0DeepSeek V3 0324
¥2.0 / ¥8.0DeepSeek V3 (Dec)
¥2.0 / ¥8.0DeepSeek R1 Distill Llama 70B
¥4.0 / ¥16.0LAST UPDATED: 2026-10-12