DeepSeek-R1-0528 显存与内存需求

具体的模型版本，带来源的结构和权重文件信息。 GGUF Q4_K_M (Unsloth)：8,192 个 token 下约 377.79 GiB–443.71 GiB，1 条序列，FP16 KV，llama.cpp。

当前结果是估算，不是对应环境的实测。未核实有文件依据的权重

发布方与出处

提供方DeepSeek
Hugging Facedeepseek-ai/DeepSeek-R1-0528
Revision4236a6af538feda4548eca9ab308586007567f52
总参数量684,531,386,000
激活参数量37,000,000,000
层数61
查询头128
KV 头128 (mla)
头维度56
已记录的上下文能力163,840
完整度partial · partial

该 MoE 记录总参数 684,531,386,000、激活参数 37,000,000,000。激活参数的计算量不会当成权重驻留内存。完整驻留权重估算使用总参数量，包括未激活的 MoE 专家。

这次估算里的内存分层

文件字节、原始 KV、引擎额外占用和总需求是分开的。这不是一次运行时轨迹。

Weight artifact: 377.13 GiB (404,941,900,832 bytes)
Raw KV cache: 0.54 GiB (575,668,224 bytes)
Engine extras / workspace: 0.13 GiB–66.04 GiB
Total estimated demand: 377.79 GiB–443.71 GiB

计入的文件

量化	计入的文件	字节	运行引擎	证据
FP8 safetensors fp8 计算器: FP8 safetensors	163 个 safetensors 分片选定字节合计：1,369,062,772,000 (1,275.04 GiB / 1,369.06 GB) 第一个分片：model-00001-of-000163.safetensors 最后一个分片：model-00163-of-000163.safetensors 查看全部 163 个分片文件 `model-00001-of-000163.safetensors` `model-00002-of-000163.safetensors` `model-00003-of-000163.safetensors` `model-00004-of-000163.safetensors` `model-00005-of-000163.safetensors` `model-00006-of-000163.safetensors` `model-00007-of-000163.safetensors` `model-00008-of-000163.safetensors` `model-00009-of-000163.safetensors` `model-00010-of-000163.safetensors` `model-00011-of-000163.safetensors` `model-00012-of-000163.safetensors` `model-00013-of-000163.safetensors` `model-00014-of-000163.safetensors` `model-00015-of-000163.safetensors` `model-00016-of-000163.safetensors` `model-00017-of-000163.safetensors` `model-00018-of-000163.safetensors` `model-00019-of-000163.safetensors` `model-00020-of-000163.safetensors` `model-00021-of-000163.safetensors` `model-00022-of-000163.safetensors` `model-00023-of-000163.safetensors` `model-00024-of-000163.safetensors` `model-00025-of-000163.safetensors` `model-00026-of-000163.safetensors` `model-00027-of-000163.safetensors` `model-00028-of-000163.safetensors` `model-00029-of-000163.safetensors` `model-00030-of-000163.safetensors` `model-00031-of-000163.safetensors` `model-00032-of-000163.safetensors` `model-00033-of-000163.safetensors` `model-00034-of-000163.safetensors` `model-00035-of-000163.safetensors` `model-00036-of-000163.safetensors` `model-00037-of-000163.safetensors` `model-00038-of-000163.safetensors` `model-00039-of-000163.safetensors` `model-00040-of-000163.safetensors` `model-00041-of-000163.safetensors` `model-00042-of-000163.safetensors` `model-00043-of-000163.safetensors` `model-00044-of-000163.safetensors` `model-00045-of-000163.safetensors` `model-00046-of-000163.safetensors` `model-00047-of-000163.safetensors` `model-00048-of-000163.safetensors` `model-00049-of-000163.safetensors` `model-00050-of-000163.safetensors` `model-00051-of-000163.safetensors` `model-00052-of-000163.safetensors` `model-00053-of-000163.safetensors` `model-00054-of-000163.safetensors` `model-00055-of-000163.safetensors` `model-00056-of-000163.safetensors` `model-00057-of-000163.safetensors` `model-00058-of-000163.safetensors` `model-00059-of-000163.safetensors` `model-00060-of-000163.safetensors` `model-00061-of-000163.safetensors` `model-00062-of-000163.safetensors` `model-00063-of-000163.safetensors` `model-00064-of-000163.safetensors` `model-00065-of-000163.safetensors` `model-00066-of-000163.safetensors` `model-00067-of-000163.safetensors` `model-00068-of-000163.safetensors` `model-00069-of-000163.safetensors` `model-00070-of-000163.safetensors` `model-00071-of-000163.safetensors` `model-00072-of-000163.safetensors` `model-00073-of-000163.safetensors` `model-00074-of-000163.safetensors` `model-00075-of-000163.safetensors` `model-00076-of-000163.safetensors` `model-00077-of-000163.safetensors` `model-00078-of-000163.safetensors` `model-00079-of-000163.safetensors` `model-00080-of-000163.safetensors` `model-00081-of-000163.safetensors` `model-00082-of-000163.safetensors` `model-00083-of-000163.safetensors` `model-00084-of-000163.safetensors` `model-00085-of-000163.safetensors` `model-00086-of-000163.safetensors` `model-00087-of-000163.safetensors` `model-00088-of-000163.safetensors` `model-00089-of-000163.safetensors` `model-00090-of-000163.safetensors` `model-00091-of-000163.safetensors` `model-00092-of-000163.safetensors` `model-00093-of-000163.safetensors` `model-00094-of-000163.safetensors` `model-00095-of-000163.safetensors` `model-00096-of-000163.safetensors` `model-00097-of-000163.safetensors` `model-00098-of-000163.safetensors` `model-00099-of-000163.safetensors` `model-00100-of-000163.safetensors` `model-00101-of-000163.safetensors` `model-00102-of-000163.safetensors` `model-00103-of-000163.safetensors` `model-00104-of-000163.safetensors` `model-00105-of-000163.safetensors` `model-00106-of-000163.safetensors` `model-00107-of-000163.safetensors` `model-00108-of-000163.safetensors` `model-00109-of-000163.safetensors` `model-00110-of-000163.safetensors` `model-00111-of-000163.safetensors` `model-00112-of-000163.safetensors` `model-00113-of-000163.safetensors` `model-00114-of-000163.safetensors` `model-00115-of-000163.safetensors` `model-00116-of-000163.safetensors` `model-00117-of-000163.safetensors` `model-00118-of-000163.safetensors` `model-00119-of-000163.safetensors` `model-00120-of-000163.safetensors` `model-00121-of-000163.safetensors` `model-00122-of-000163.safetensors` `model-00123-of-000163.safetensors` `model-00124-of-000163.safetensors` `model-00125-of-000163.safetensors` `model-00126-of-000163.safetensors` `model-00127-of-000163.safetensors` `model-00128-of-000163.safetensors` `model-00129-of-000163.safetensors` `model-00130-of-000163.safetensors` `model-00131-of-000163.safetensors` `model-00132-of-000163.safetensors` `model-00133-of-000163.safetensors` `model-00134-of-000163.safetensors` `model-00135-of-000163.safetensors` `model-00136-of-000163.safetensors` `model-00137-of-000163.safetensors` `model-00138-of-000163.safetensors` `model-00139-of-000163.safetensors` `model-00140-of-000163.safetensors` `model-00141-of-000163.safetensors` `model-00142-of-000163.safetensors` `model-00143-of-000163.safetensors` `model-00144-of-000163.safetensors` `model-00145-of-000163.safetensors` `model-00146-of-000163.safetensors` `model-00147-of-000163.safetensors` `model-00148-of-000163.safetensors` `model-00149-of-000163.safetensors` `model-00150-of-000163.safetensors` `model-00151-of-000163.safetensors` `model-00152-of-000163.safetensors` `model-00153-of-000163.safetensors` `model-00154-of-000163.safetensors` `model-00155-of-000163.safetensors` `model-00156-of-000163.safetensors` `model-00157-of-000163.safetensors` `model-00158-of-000163.safetensors` `model-00159-of-000163.safetensors` `model-00160-of-000163.safetensors` `model-00161-of-000163.safetensors` `model-00162-of-000163.safetensors` `model-00163-of-000163.safetensors`	1,369,062,772,000 (1,275.04 GiB / 1,369.06 GB)	`vllm, sglang, unspecified`	`file-metadata`
GGUF Q4_K_M (Unsloth) gguf-q4_k_m 计算器: GGUF Q4_K_M (Unsloth)	9 个 safetensors 分片选定字节合计：404,941,900,832 (377.13 GiB / 404.94 GB) 第一个分片：Q4_K_M/DeepSeek-R1-0528-Q4_K_M-00001-of-00009.gguf 最后一个分片：Q4_K_M/DeepSeek-R1-0528-Q4_K_M-00009-of-00009.gguf 查看全部 9 个分片文件 `Q4_K_M/DeepSeek-R1-0528-Q4_K_M-00001-of-00009.gguf` `Q4_K_M/DeepSeek-R1-0528-Q4_K_M-00002-of-00009.gguf` `Q4_K_M/DeepSeek-R1-0528-Q4_K_M-00003-of-00009.gguf` `Q4_K_M/DeepSeek-R1-0528-Q4_K_M-00004-of-00009.gguf` `Q4_K_M/DeepSeek-R1-0528-Q4_K_M-00005-of-00009.gguf` `Q4_K_M/DeepSeek-R1-0528-Q4_K_M-00006-of-00009.gguf` `Q4_K_M/DeepSeek-R1-0528-Q4_K_M-00007-of-00009.gguf` `Q4_K_M/DeepSeek-R1-0528-Q4_K_M-00008-of-00009.gguf` `Q4_K_M/DeepSeek-R1-0528-Q4_K_M-00009-of-00009.gguf`	404,941,900,832 (377.13 GiB / 404.94 GB)	`llama.cpp, ollama, unspecified`	`file-metadata`
GGUF Q5_K_M (Unsloth) gguf-q5_k_m 计算器: GGUF Q5_K_M (Unsloth)	10 个 safetensors 分片选定字节合计：475,780,344,992 (443.10 GiB / 475.78 GB) 第一个分片：Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00001-of-00010.gguf 最后一个分片：Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00010-of-00010.gguf 查看全部 10 个分片文件 `Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00001-of-00010.gguf` `Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00002-of-00010.gguf` `Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00003-of-00010.gguf` `Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00004-of-00010.gguf` `Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00005-of-00010.gguf` `Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00006-of-00010.gguf` `Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00007-of-00010.gguf` `Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00008-of-00010.gguf` `Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00009-of-00010.gguf` `Q5_K_M/DeepSeek-R1-0528-Q5_K_M-00010-of-00010.gguf`	475,780,344,992 (443.10 GiB / 475.78 GB)	`llama.cpp, ollama, unspecified`	`file-metadata`
GGUF Q6_K (Unsloth) gguf-q6_k 计算器: GGUF Q6_K (Unsloth)	12 个 safetensors 分片选定字节合计：551,046,192,128 (513.20 GiB / 551.05 GB) 第一个分片：Q6_K/DeepSeek-R1-0528-Q6_K-00001-of-00012.gguf 最后一个分片：Q6_K/DeepSeek-R1-0528-Q6_K-00012-of-00012.gguf 查看全部 12 个分片文件 `Q6_K/DeepSeek-R1-0528-Q6_K-00001-of-00012.gguf` `Q6_K/DeepSeek-R1-0528-Q6_K-00002-of-00012.gguf` `Q6_K/DeepSeek-R1-0528-Q6_K-00003-of-00012.gguf` `Q6_K/DeepSeek-R1-0528-Q6_K-00004-of-00012.gguf` `Q6_K/DeepSeek-R1-0528-Q6_K-00005-of-00012.gguf` `Q6_K/DeepSeek-R1-0528-Q6_K-00006-of-00012.gguf` `Q6_K/DeepSeek-R1-0528-Q6_K-00007-of-00012.gguf` `Q6_K/DeepSeek-R1-0528-Q6_K-00008-of-00012.gguf` `Q6_K/DeepSeek-R1-0528-Q6_K-00009-of-00012.gguf` `Q6_K/DeepSeek-R1-0528-Q6_K-00010-of-00012.gguf` `Q6_K/DeepSeek-R1-0528-Q6_K-00011-of-00012.gguf` `Q6_K/DeepSeek-R1-0528-Q6_K-00012-of-00012.gguf`	551,046,192,128 (513.20 GiB / 551.05 GB)	`llama.cpp, ollama, unspecified`	`file-metadata`
GGUF Q8_0 (Unsloth) gguf-q8_0 计算器: GGUF Q8_0 (Unsloth)	15 个 safetensors 分片选定字节合计：713,286,522,336 (664.30 GiB / 713.29 GB) 第一个分片：Q8_0/DeepSeek-R1-0528-Q8_0-00001-of-00015.gguf 最后一个分片：Q8_0/DeepSeek-R1-0528-Q8_0-00015-of-00015.gguf 查看全部 15 个分片文件 `Q8_0/DeepSeek-R1-0528-Q8_0-00001-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00002-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00003-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00004-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00005-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00006-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00007-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00008-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00009-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00010-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00011-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00012-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00013-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00014-of-00015.gguf` `Q8_0/DeepSeek-R1-0528-Q8_0-00015-of-00015.gguf`	713,286,522,336 (664.30 GiB / 713.29 GB)	`llama.cpp, ollama, unspecified`	`file-metadata`
INT4 theoretical	没有已发布的文件	342,265,693,000 (理论行不是可下载的发布文件。)	—	theoretical

按量化与上下文查看显存

并发序列数为 1。KV 默认 FP16。区间包含所列引擎已记录的运行时额外占用。这些数字与交互式计算器使用同一套核心。

量化	2,048 tokens	8,192 tokens	32,768 tokens
FP8 safetensors	1,275.80 GiB–1,469.35 GiB	1,276.20 GiB–1,469.90 GiB	1,277.81 GiB–1,472.09 GiB
GGUF Q4_K_M (Unsloth)	377.39 GiB–443.13 GiB	377.79 GiB–443.71 GiB	379.40 GiB–445.98 GiB
GGUF Q5_K_M (Unsloth)	443.36 GiB–520.52 GiB	443.77 GiB–521.10 GiB	445.37 GiB–523.37 GiB
GGUF Q6_K (Unsloth)	513.46 GiB–602.74 GiB	513.86 GiB–603.32 GiB	515.47 GiB–605.60 GiB
GGUF Q8_0 (Unsloth)	664.56 GiB–779.98 GiB	664.96 GiB–780.56 GiB	666.57 GiB–782.83 GiB
INT4 theoretical 理论行不是可下载的发布文件。	319.02 GiB–411.32 GiB	319.42 GiB–411.90 GiB	321.03 GiB–414.17 GiB

KV Cache（原始数据）

用这个模型打开计算器

数据来源

来源
来源 — revision 4236a6af538feda4548eca9ab308586007567f52
来源 — metadata.total_size=1369062772000
来源

影响结论的假设

请求的上下文在记录的原生窗口 163840 token 以内。
完整驻留权重估算使用总参数量，包括未激活的 MoE 专家。
该 MoE 记录总参数 684531386000、激活参数 37000000000。激活参数的计算量不会当成权重驻留内存。
权重字节只计入选定的权重文件，不会把仓库里所有量化版本加在一起。
下载文件大小不是峰值 GPU 内存。运行时布局、分配器填充和 KV 是额外占用。
权重字节来自 9 个已发布的 safetensors 分片，合计 404941900832 字节。
存在 MTP / draft 权重；未启用时不计入。
上下文预算是每个序列缓存的 token 数（输入加上预留的生成长度）。

计算器

正在加载已发布的目录快照…

先选择模型，再检查设备。

模型

显示历史模型

量化

设备类型

硬件

工作负载

上下文长度主要影响 KV Cache 内存。

高级设置

运行时

这次估算使用的引擎和 KV 缓存类型。

运行引擎KV Cache 精度循环 / 卷积状态精度包含理论估算（专家）筛选无匹配时允许替代

工作负载

开销场景，以及是否计入预填充工作区。

运行时开销情景计入更重的预填充工作区

拓扑

模型如何切到多张设备上。各设备的显存不会相加。

部署方式组内切分设备数量

副本数是高级部署选项。数量指节点或 GPU，不是副本。

副本组数每个副本的设备数

预算

系统保留和安全余量。数字框清空时保留上一次的值。

系统 / 显示预留

GiB（2^30 字节）

预算已经扣除系统占用额外安全余量

自定义每参数有效 bit 数显示单位

自定义模型

导入配置或手填结构。点导入之前不会去扫描仓库。

导入本地 config.json

仅在浏览器中解析，不会上传文件。

公开的 Hugging Face 仓库Revision

只允许 huggingface.co 上的 config.json，不使用 token，不读取私有仓库。

总参数量激活参数量层数查询头KV 头头维度

搜索或选择目录中的模型，查看本地大模型的显存需求。

搜索或选择目录中的模型，查看本地大模型的显存需求。

DeepSeek-R1-0528 显存与内存需求

发布方与出处

这次估算里的内存分层

计入的文件

按量化与上下文查看显存

KV Cache（原始数据）

相关页面

数据来源

影响结论的假设

计算器