Overview

12,500 道竞赛数学题（含 7 个学科：代数、几何、数论、微积分、概率等，5 个难度级别），评估模型的复杂数学推理能力。

Metrics

Metric	Unit	Direction
accuracy	%	↑ Higher is better

Sources

Model Score Ranking

#	Model	Vendor	Score
1	Qwen2.5 72B	alibaba	83.1
2	Grok-2	xai	76.8
3	GPT-4o	openai	76.6
4	Claude 3.5 Sonnet (2024-10-22)	anthropic	75.1
5	Llama 3.1 405B	meta	73.8
6	Llama 3.3 70B	meta	73.8
7	Qwen2 72B	alibaba	71.9
8	o1 Preview	openai	71.5
9	Gemini 2.0 Flash	google	71.3
10	Claude 3.5 Sonnet	anthropic	71.1
11	Mistral Large 2	mistral	71
12	GPT-4o (2024-08-06)	openai	68.2
13	GPT-4o (2024-05-13)	openai	67.1
14	Microsoft WizardMath 7B v1	other	65.8
15	Hermes 3 Llama 3.1 405B	other	65.7
16	DeepSeek V3	deepseek	61.6
17	Mathstral 7B	mistral	61
18	Jamba 1.5 Large	other	60
19	Yi Vision	other	58.7
20	Gemini 1.5 Pro	google	58.5
21	Claude 3 Opus (2024-02-29)	anthropic	58.2
22	Gemini 1.0 Pro	google	58.2
23	Gemini 1.0 Ultra	google	57.1
24	Gemini 2.0 Flash Thinking	google	56.6
25	Gemini 1.5 Pro 002	google	55.9
26	Yi Large	other	55.9
27	o1	openai	55.7
28	Mistral Medium	mistral	54.9
29	Qwen2 57B	alibaba	54.8
30	Hermes 3 Llama 3.1 70B	other	54.4
31	Qwen1.5 110B	alibaba	54.1
32	DeepSeek Math 7B	deepseek	53.8
33	GLM-4 Plus	other	53.3
34	Gemini 1.5 Flash	google	53.2
35	Claude 3.5 Haiku	anthropic	53
36	Llama 3.2 90B Vision	meta	52.5
37	Claude 3 Haiku	anthropic	52.2
38	GPT-4 1106 Preview	openai	52.1
39	GPT-4o mini	openai	51.8
40	Sonar Large	other	51.6
41	GPT-4 Turbo	openai	50.9
42	Sonar Reasoning	other	50.7
43	Yi 1.5 34B	other	50.4
44	DBRX Instruct	other	50.1
45	Llama 3 70B	meta	50.1
46	GPT-4 Vision Preview	openai	49.7
47	DeepSeek V2 Chat	deepseek	49.3
48	Mistral Large	mistral	49.3
49	Command Nightly	cohere	49.2
50	Jamba Instruct	other	49
51	GPT-4 Vision	openai	48.3
52	Jamba 1.5	other	48.1
53	Llemma 7B	other	47.5
54	Code Llama 7B	meta	47.1
55	Command R (08-2024)	cohere	47.1
56	Sonar Huge	other	46.8
57	Claude 3 Sonnet (2024-02-29)	anthropic	46.4
58	Qwen1.5 32B	alibaba	46.3
59	Qwen1.5 14B	alibaba	46.2
60	StarCoder2 15B	other	46.2
61	Mixtral 8x22B	mistral	46
62	Mistral Small	mistral	45.8
63	StarChat2 15B v0.1	other	45.3
64	Grok-2 Mini	xai	45.1
65	Nous Hermes 2 Yi 34B	other	44.9
66	Falcon 180B	other	44.8
67	Claude 3 Haiku (2024-03-07)	anthropic	44.7
68	Code Llama 34B	meta	44.7
69	Yi 1.5 6B	other	44.5
70	o1 mini	openai	44.2
71	Command R	cohere	43.9
72	GPT-4	openai	43.9
73	Mistral Nemo	mistral	43.6
74	Phi-3 Small	other	43.5
75	Command R+ (08-2024)	cohere	43.3
76	Sonar Small	other	43.3
77	OLMo 7B SFT	other	43.1
78	Gemini 1.5 Flash-8B 002	google	43
79	Claude 3 Opus	anthropic	42.7
80	Nous Hermes 2 Mixtral 8x7B	other	42.6
81	Jamba 1.5 Mini	other	42.5
82	Claude 3 Sonnet	anthropic	42.4
83	GPT-4 0125 Preview	openai	42.3
84	Orca 2 13B	other	42.3
85	Yi Large Turbo	other	42.2
86	Qwen2.5 7B	alibaba	41.9
87	Grok-2 Vision	xai	41.3
88	Qwen2 7B	alibaba	40.7
89	Command R7B	cohere	40.6
90	GLM-4 Air	other	40.6
91	Microsoft WizardCoder Python 34B	other	40.6
92	Gemini 1.5 Flash 002	google	40.5
93	NVIDIA Llama 3.1 Nemotron 70B	other	40.3
94	GPT-4 32K	openai	40.1
95	Llama 3.1 Nemotron 70B	meta	40
96	Llama 3 8B	meta	40
97	Gemini 1.0 Flash	google	39.9
98	Ministral 8B	mistral	39.9
99	Microsoft WizardLM 2 8x22B	other	39.8
100	Phi-4	other	39.6
101	Qwen2.5 32B	alibaba	39.5
102	GLM-4 Flash	other	39.4
103	DeepSeek LLM 67B	deepseek	39.3
104	OLMo 2 1124 7B	other	39.3
105	Mistral Small 3	mistral	39.2
106	Qwen2.5 14B	alibaba	38.8
107	Llama 3.1 70B	meta	38.5
108	DeepSeek Coder V2	deepseek	38.2
109	Phi-3.5 MoE	other	38.2
110	OLMo 1.7 7B	other	38.1
111	Mixtral 8x7B	mistral	38
112	StableLM 2 12B	other	37.9
113	Code Llama 13B	meta	37.8
114	Nous Hermes 2 Solar 10.7B	other	37.7
115	Qwen1.5 72B	alibaba	37.6
116	StarCoder2 3B	other	37.4
117	WizardLM Team WizardLM 2 8x22B	other	37.4
118	OLMo 7B	other	37.1
119	DBRX Base	other	36
120	Zephyr ORPO 141B Alpha	other	36
121	Command R+	cohere	35.6
122	Gemma 2 27B	google	35.5
123	DeepSeek V2	deepseek	35.2
124	Gemini 1.5 Flash-8B	google	35
125	OLMo 7B Instruct	other	34.7
126	Code Bison	google	34.4
127	Phi-3 Medium	other	34.3
128	Yi 1.5 9B	other	34.1
129	Code Llama 70B	meta	34
130	OpenChat 3.6 8B	other	33.7
131	GLM-4V 9B	other	33.3
132	DeepSeek Coder 7B	deepseek	32.2
133	DeepSeek Coder 33B	deepseek	32.1
134	ChatGLM3 6B	other	31.8
135	Zephyr 7B Beta	other	31.3
136	Llama 3.1 8B	meta	31.1
137	Argilla Notus 7B v1	other	31
138	StableCode 3B	other	31
139	Microsoft WizardLM 2 7B	other	30.5
140	StarCoder2 7B	other	30.5
141	GPT-3.5 Turbo 16K	openai	30.4
142	Llama 3.2 11B Vision	meta	29.9
143	PaLM 2	google	29.9
144	Yi 6B	other	29.9
145	Codestral	mistral	29.4
146	GPT-3.5	openai	29.4
147	Gemma 2 9B	google	28.9
148	Codestral Mamba	mistral	28.1
149	Flan-T5 XL	other	27.8
150	Gemma 2B	google	27.5
151	Jurassic-2 Ultra	other	27.5
152	Phi-1.5	other	27.2
153	Mistral 7B v0.3	mistral	26.7
154	GLM-4 9B Chat	other	26.6
155	Mistral 7B v0.1	mistral	26.5
156	StableLM 2 1.6B	other	26.4
157	Llama Guard 2 8B	meta	26.3
158	Llama 2 13B	meta	25.9
159	Mistral 7B v0.2	mistral	25.9
160	Gemma 7B	google	25.7
161	Capybara 1.5B	other	25.6
162	Hermes 3 Llama 3.1 8B	other	25.5
163	Qwen2 1.5B	alibaba	25.4
164	Phi-1	other	25.2
165	Phi-3 Vision	other	24.3
166	Command Light	cohere	24.1
167	Grok Vision Beta	xai	24
168	Baichuan2 13B Chat	other	23.4
169	Zephyr 7B Alpha	other	23.2
170	Nous Capybara 34B	other	23.1
171	Claude Instant 1	anthropic	22.5
172	Llama 2 70B	meta	22.4
173	Flan-UL2	other	22
174	Claude 2	anthropic	21.9
175	Flan-T5 XXL	other	21.4
176	Falcon 40B	other	21.3
177	MPT 30B	other	21.2
178	Open-Platypus	other	21
179	Phi-2	other	20.5
180	Yi 34B	other	20.5
181	Jurassic-2 Mid	other	20.3
182	Chat Bison	google	19.8
183	Nous Capybara 7B	other	19.7
184	Llama 2 7B	meta	19.4
185	Grok Beta	xai	19.1
186	Baichuan2 7B Chat	other	18.8
187	MPT 7B	other	18.4
188	Phi-3 Mini	other	17.9
189	GPT-3.5 Turbo	openai	17.8
190	Qwen2.5 3B	alibaba	17
191	Mistral Tiny	mistral	16
192	Claude 2.1	anthropic	15.5
193	Qwen2.5 0.5B	alibaba	15.4
194	TigerBot 70B Chat	other	15.4
195	OpenChat 3.5 1210	other	15.2
196	StableLM Zephyr 3B	other	14.9
197	Text Bison	google	12.1
198	StableLM 3 4B	other	11.5
199	Llama 3.2 1B	meta	11.4
200	Ministral 3B	mistral	10.1
201	Llama 3.2 3B	meta	10
202	Phi-3.5 Mini	other	9.6
203	Qwen2.5 1.5B	alibaba	9.5
204	Llama Guard 3 8B	meta	8.4
205	Embed English v3	cohere	0

MATH

설명

12,500 道竞赛数学题（含 7 个学科：代数、几何、数论、微积分、概率等，5 个难度级别），评估模型的复杂数学推理能力。

핵심 사양

카테고리	라이선스	최종 업데이트
math	MIT	2021-01-01

벤치마크

단위
%

소스

공식 URL

https://github.com/hendrycks/math