Deepseek V4.1 Flash

DeepSeek2026-09-10● 开源 · MIT🤗 deepseek-ai/DeepSeek-V4.1-Flash129.5万4,415

DeepSeek V4.1 Flash 是 DeepSeek 推出的稀疏混合专家(MoE)模型,也是该公司首款采用「因果编码器-解码器」(Causal Encoder-Decoder, CED)架构的模型。输入端激活 80 亿参数、输出端激活 160 亿参数,总参数规模 5520 亿,这种非对称设计让单 token 计算量相对总参数量保持较低。图像理解能力原生集成在架构中,视觉与文本嵌入从预训练一开始就联合训练,而不是像此前的实验性版本 V4 Flash Vision Exp 那样后期附加。

多服务商价格对比数据来自 OpenRouter

32 个服务商1 USD ≈ 6.7058 CNY · 更新于 2026-10-13

服务商	量化精度	上下文	输入/M	输出/M	缓存读取/M	近1天可用率
O OpenInference	fp4	1.05M	¥0.30	¥5.97	¥0.04	98.7%
Relace	—	1.05M	¥0.34	¥4.02	¥0.05	98.0%
Morph	fp8	1.05M	¥0.40	¥4.36	¥0.06	100.0%
I InferenceNet	fp8	1.04M	¥0.52	¥3.69	¥0.07	100.0%
D Decart	fp4	1.05M	¥0.60	¥1.21	¥0.12	100.0%
S Sail Research	fp4	1.05M	¥0.60	¥2.01	¥0.04	99.9%
I Ionstream	—	1.05M	¥0.67	¥5.03	¥0.07	97.8%
W Wafer	—	1.05M	¥0.87	¥6.64	¥0.23	100.0%
DeepInfra限时 -30%	fp8	1.05M	¥0.94	¥2.82	¥0.03	100.0%
StreamLake限时 -51%	fp8	1.02M	¥0.99	¥3.94	¥0.02	100.0%
DigitalOcean	—	1.05M	¥1.11	¥4.43	¥0.04	100.0%
G GMICloud限时 -40%	fp8	1.05M	¥1.21	¥4.83	¥0.02	100.0%
Novita限时 -35%	fp8	1.05M	¥1.31	¥5.23	¥0.03	100.0%
C CoreWeave	fp8	1.05M	¥1.34	¥4.36	¥0.20	100.0%
P Phala限时 -30%	—	1.05M	¥1.41	¥5.63	¥0.03	100.0%
M Makora	fp8	1.05M	¥1.81	¥7.71	¥0.06	100.0%
Crusoe	fp8	1.05M	¥1.94	¥8.05	¥0.05	99.8%
Baidu	fp8	1.05M	¥2.01	¥8.05	¥0.04	97.6%
Cloudflare	—	1.05M	¥2.01	¥8.05	¥0.04	100.0%
AtlasCloud	fp8	1.05M	¥2.01	¥8.05	¥0.20	100.0%
BaseTen	fp8	1.05M	¥2.01	¥8.05	¥0.05	100.0%
Alibaba 时段定价	—	1.00M	¥2.01	¥8.05	¥0.20	98.1%
Together	—	1.05M	¥2.01	¥8.05	¥0.04	100.0%
S SiliconFlow	fp8	1.05M	¥2.01	¥8.05	¥0.04	99.7%
Modal	—	1.05M	¥2.01	¥8.05	¥0.20	100.0%
BaseTen	fp8	1.05M	¥2.01	¥8.05	¥0.05	100.0%
Parasail	fp8	1.05M	¥2.01	¥8.05	¥0.04	99.9%
Fireworks	—	1.05M	¥2.01	¥8.05	¥0.04	100.0%
Venice	fp8	1.00M	¥2.01	¥8.05	¥0.05	100.0%
DeepSeek	—	1.05M	¥2.01	¥8.05	¥0.04	100.0%
Fireworks (US)	—	1.05M	¥3.02	¥12.07	¥0.06	100.0%
BaseTen	fp32	1.05M	¥4.02	¥16.09	¥0.09	99.2%

O

OpenInferencefp4

¥0.30

输入/M

¥5.97

输出/M

¥0.04

缓存/M

98.7%

可用率

Relace

¥0.34

输入/M

¥4.02

输出/M

¥0.05

缓存/M

98.0%

可用率

Morphfp8

¥0.40

输入/M

¥4.36

输出/M

¥0.06

缓存/M

100.0%

可用率

I

InferenceNetfp8

¥0.52

输入/M

¥3.69

输出/M

¥0.07

缓存/M

100.0%

可用率

D

Decartfp4

¥0.60

输入/M

¥1.21

输出/M

¥0.12

缓存/M

100.0%

可用率

S

Sail Researchfp4

¥0.60

输入/M

¥2.01

输出/M

¥0.04

缓存/M

99.9%

可用率

I

Ionstream

¥0.67

输入/M

¥5.03

输出/M

¥0.07

缓存/M

97.8%

可用率

W

Wafer

¥0.87

输入/M

¥6.64

输出/M

¥0.23

缓存/M

100.0%

可用率

DeepInfrafp8限时 -30%

¥0.94

输入/M

¥2.82

输出/M

¥0.03

缓存/M

100.0%

可用率

StreamLakefp8限时 -51%

¥0.99

输入/M

¥3.94

输出/M

¥0.02

缓存/M

100.0%

可用率

DigitalOcean

¥1.11

输入/M

¥4.43

输出/M

¥0.04

缓存/M

100.0%

可用率

G

GMICloudfp8限时 -40%

¥1.21

输入/M

¥4.83

输出/M

¥0.02

缓存/M

100.0%

可用率

Novitafp8限时 -35%

¥1.31

输入/M

¥5.23

输出/M

¥0.03

缓存/M

100.0%

可用率

C

CoreWeavefp8

¥1.34

输入/M

¥4.36

输出/M

¥0.20

缓存/M

100.0%

可用率

P

Phala限时 -30%

¥1.41

输入/M

¥5.63

输出/M

¥0.03

缓存/M

100.0%

可用率

M

Makorafp8

¥1.81

输入/M

¥7.71

输出/M

¥0.06

缓存/M

100.0%

可用率

Crusoefp8

¥1.94

输入/M

¥8.05

输出/M

¥0.05

缓存/M

99.8%

可用率

Baidufp8

¥2.01

输入/M

¥8.05

输出/M

¥0.04

缓存/M

97.6%

可用率

Cloudflare

¥2.01

输入/M

¥8.05

输出/M

¥0.04

缓存/M

100.0%

可用率

AtlasCloudfp8

¥2.01

输入/M

¥8.05

输出/M

¥0.20

缓存/M

100.0%

可用率

BaseTenfp8

¥2.01

输入/M

¥8.05

输出/M

¥0.05

缓存/M

100.0%

可用率

Alibaba 时段定价

¥2.01

输入/M

¥8.05

输出/M

¥0.20

缓存/M

98.1%

可用率

Together

¥2.01

输入/M

¥8.05

输出/M

¥0.04

缓存/M

100.0%

可用率

S

SiliconFlowfp8

¥2.01

输入/M

¥8.05

输出/M

¥0.04

缓存/M

99.7%

可用率

Modal

¥2.01

输入/M

¥8.05

输出/M

¥0.20

缓存/M

100.0%

可用率

BaseTenfp8

¥2.01

输入/M

¥8.05

输出/M

¥0.05

缓存/M

100.0%

可用率

Parasailfp8

¥2.01

输入/M

¥8.05

输出/M

¥0.04

缓存/M

99.9%

可用率

Fireworks

¥2.01

输入/M

¥8.05

输出/M

¥0.04

缓存/M

100.0%

可用率

Venicefp8

¥2.01

输入/M

¥8.05

输出/M

¥0.05

缓存/M

100.0%

可用率

DeepSeek

¥2.01

输入/M

¥8.05

输出/M

¥0.04

缓存/M

100.0%

可用率

Fireworks (US)

¥3.02

输入/M

¥12.07

输出/M

¥0.06

缓存/M

100.0%

可用率

BaseTenfp32

¥4.02

输入/M

¥16.09

输出/M

¥0.09

缓存/M

99.2%

可用率