新一代大模型架构 超越Transformer

为您介绍 RWKV

RWKV 结合了 RNN 和 Transformer 的所有优点,既可高效并行推理训练,亦可高效串行推理训练

下载 RWKV 模型
您可下载 RWKV-7"Goose"(最新最强架构)模型,和 RWKV-6 "Finch" 模型
RWKV-7 G1 "GooseOne" 0.1B 20250307
RWKV-7 G1 "GooseOne"
0.1B 20250307
RWKV7-G1 "GooseOne" 0.4B 20250324
RWKV7-G1 "GooseOne"
0.4B 20250324
RWKV-7 World-v3 1.5B 20250127
RWKV-7 World-v3
1.5B 20250127
RWKV-7 World-v3 2.9B 20250211
RWKV-7 World-v3
2.9B 20250211
RWKV-6 World-v3 7B 20241112
RWKV-6 World-v3
7B 20241112
RWKV 的测评成绩
"英文与多语言基准测试"涵盖多种评测类型(如 LAMBADA、PIQA 和 WinoGrande),评估模型在语言理解、推理和知识泛化能力方面的表现。
英文和多语言测评
Uncheatable 3B
Uncheatable 0.4B
lm-evaluation-harness
RWKV-7 "Goose" World-v3
Llama-3.2-3B (+emb = 3.61b)
Qwen2.5-3B (+emb = 3.40b)
stablelm-3b-4e1t
RWKV-6 'Finch' World-v2.1
Params
English
LAMBADA
PIQA
StoryCloze16
Hellaswag
WinoGrande
arc_challenge
arc_easy
headQA_en
openbookQA
sciq
ReCoRD
MultiLang
xLBD
xSC
xWG
xCOPA
2.90
71.1
73.0
79.6
79.4
76.5
72.4
50.5
79.5
42.3
46.6
94.7
87.7
62.3
47.2
63.3
76.0
62.8
3.61
68.7
70.2
76.5
78.0
73.6
69.5
46.3
74.2
39.8
42.8
95.4
89.6
57.3
39.6
59.2
71.8
58.6
3.40
68.6
66.9
78.4
77.2
73.6
68.1
47.2
77.4
40.7
43.0
95.3
87.1
57.0
37.8
58.0
73.2
58.9
2.80
67.5
70.6
79.6
76.9
73.8
66.5
39.9
72.3
39.0
40.4
94.6
88.9
54.6
39.2
56.1
66.7
56.4
3.10
66.0
71.4
75.8
78.6
68.2
65.4
39.0
71.1
36.6
40.6
92.9
86.3
59.2
44.6
59.9
71.8
60.6
lm-evaluation-harness
RWKV-7 "Goose" World-v3
SmolLM2-1.7B (+emb = 1.81b)
Qwen2.5-1.5B (+emb = 1.78b)
stablelm-2-1_6b
RWKV-6 'Finch' World-v2.1
Llama-3.2-1B (+emb = 1.50b)
Params
English
LAMBADA
PIQA
StoryCloze16
Hellaswag
WinoGrande
arc_challenge
arc_easy
headQA_en
openbookQA
sciq
ReCoRD
MultiLang
xLBD
xSC
xWG
xCOPA
1.52
67.7
69.4
77.2
76.3
70.8
67.9
42.5
75.8
39.0
46.4
94.4
85.3
58.6
43.0
59.7
72.2
59.7
1.81
67.9
67.4
76.7
76.0
71.5
66.7
47.0
78.0
39.8
44.6
93.4
86.3
49.2
29.3
52.0
62.5
53.1
1.78
65.5
62.2
75.8
73.8
67.9
63.5
45.3
75.4
37.6
40.2
94.6
83.9
53.8
34.1
55.9
68.1
57.1
1.64
65.1
65.8
76.3
75.8
68.9
64.3
39.9
68.6
34.3
39.8
95.3
86.8
52.3
34.8
53.8
64.4
56.3
1.60
62.2
67.3
73.9
74.9
61.1
60.5
34.2
64.2
35.7
38.6
90.3
83.7
56.2
40.9
56.7
69.2
58.1
1.50
62.1
62.2
74.6
72.4
63.6
60.5
36.2
65.7
34.0
37.4
90.9
85.9
51.8
33.0
55.2
63.7
55.4
lm-evaluation-harness
RWKV7-G1 "GooseOne" 20250324
SmolLM2-360M (+emb = 407m)
RWKV-7 "Goose" Pile
mamba-370m (+emb = 421m) Pile
Params
English
LAMBADA
PIQA
StoryCloze16
Hellaswag
WinoGrande
arc_challenge
arc_easy
headQA_en
openbookQA
sciq
ReCoRD
MultiLang
xLBD
xSC
xWG
xCOPA
0.450
60.5
59.3
72.1
70.3
57.4
59.3
34.9
68.3
34.1
40.0
91.1
78.6
52.9
34.5
55.3
65.9
55.7
0.407
59.7
53.4
71.8
68.2
56.4
59.0
37.9
70.2
34.3
37.8
90.8
77.4
43.7
18.7
49.7
54.5
51.8
0.421
55.8
57.9
69.2
67.7
48.1
56.4
27.6
56.2
32.1
32.2
85.9
80.2
47.6
29.4
50.7
57.3
52.9
0.421
54.7
55.6
69.5
66.3
46.5
55.5
27.9
55.0
32.3
30.8
84.9
77.0
47.2
28.5
50.5
57.3
52.4
lm-evaluation-harness
RWKV7-G1 "GooseOne" 20250307
SmolLM2-135M (+emb = 163m)
RWKV-7 "Goose" Pile
mamba-130m (+emb = 168m) Pile
Params
English
LAMBADA
PIQA
StoryCloze16
Hellaswag
WinoGrande
arc_challenge
arc_easy
headQA_en
openbookQA
sciq
ReCoRD
MultiLang
xLBD
xSC
xWG
xCOPA
0.191
52.8
49.1
68.0
64.3
42.9
53.8
28.2
56.1
30.3
32.4
85.8
70.5
47.6
27.1
51.9
58.3
53.4
0.163
53.0
42.9
68.2
63.4
43.2
53.1
29.7
64.3
31.1
33.0
83.8
70.1
41.7
12.5
49.4
52.8
52.0
0.168
49.8
45.6
65.5
61.8
36.9
52.3
24.7
47.9
29.1
30.0
81.8
71.9
43.9
21.7
49.3
52.0
52.8
0.168
48.5
44.2
64.4
60.4
35.3
52.4
24.3
48.1
28.8
28.6
78.1
68.9
43.7
20.1
49.2
53.2
52.4
我们拥有丰富的全球开发者生态
作为能效最高、表达力超越 attention,完全开源可商用的新型大模型架构,RWKV 广受开发者和爱好者们的好评。
目前 RWKV 开源社区的国内开发者超过 1 万人、海外开发者超过 8000 人、Github 上基于 RWKV 的开源项目超过 400 个
我们有众多基于 RWKV 大模型的应用案例
RWKV 作曲家
关于我们
还有更多可以深刻挖掘的信息