Global-MMLU-Lite · Open Analysis

#	Model	Effort	Score	Cost / task	Harness
1	Gemini 3.1 Pro	–	93.2%	–	–
2	Gemini 3 Flash	none	92.7%	–	–
3	Gemini 3 Pro	high	92.2%	–	–
4	Opus 4.6	max	92.2%	–	–
5	Opus 4.5	–	91.3%	–	–
6	GPT-5	high	90.7%	–	–
7	GPT-5.1	high	90.6%	–	–
8	Sonnet 4.6	max	90.5%	–	–
9	Gemini 2.5 Pro	–	90.3%	–	–
10	Qwen 3.5 397B-A17B	–	90.0%	–	–
11	GPT-5.2	medium	89.8%	–	–
12	GPT-5.1 Codex	high	89.7%	–	–
13	Grok 4	–	89.5%	–	–
14	Grok 4.20	–	89.5%	–	–
15	GPT-5.4 Mini	xhigh	89.4%	–	–
16	Sonnet 4.5	–	89.3%	–	–
17	Qwen 3.5 122B-A10B	–	89.1%	–	–
18	Gemini 2.5 Flash	–	88.8%	–	–
19	DeepSeek V3.2	–	88.4%	–	–
20	Nova 2.0 Pro Preview (Low)	low	87.6%	–	–
21	GPT-5 Mini	high	87.4%	–	–
22	Qwen 3.5 27B	–	86.8%	–	–
23	ERNIE 5.0 Thinking Preview	–	86.5%	–	–
24	Qwen 3.5 35B-A3B	–	86.4%	–	–
25	DeepSeek R1	–	86.0%	–	–
26	Grok 4 Fast	–	85.9%	–	–
27	Grok 4.1 Fast	–	85.6%	–	–
28	GLM 4.6	–	85.6%	–	–
29	Qwen 3 Max	max	85.5%	–	–
30	Qwen 3 235B-A22B	–	85.3%	–	–
31	GPT-5.1 Codex Mini	high	85.0%	–	–
32	MiMo V2 Flash	–	84.4%	–	–
33	MiniMax M2.5	–	84.2%	–	–
34	Kimi K2.5	none	84.0%	–	–
35	MiniMax M2.1	–	84.0%	–	–
36	Gemini 2.5 Flash Lite	–	83.8%	–	–
37	Haiku 4.5	–	83.4%	–	–
38	gpt-oss-120b	high	82.8%	–	–
39	DeepSeek V3.1	–	82.7%	–	–
40	Llama 4 Maverick	none	82.5%	–	–
41	Qwen 3	–	81.9%	–	–
42	GLM 5	none	81.9%	–	–
43	KAT-Coder-Pro V1	none	81.7%	–	–
44	Solar Open 100B (Reasoning)	high	81.6%	–	–
45	Magistral Medium 1.2	medium	81.6%	–	–
46	Mi:dm K 2.5 Pro Preview	–	81.4%	–	–
47	GPT-5 Nano	high	81.2%	–	–
48	Nova 2.0 Omni (Low)	low	81.0%	–	–
49	Nova 2.0 Omni (Medium)	medium	81.0%	–	–
50	Llama Nemotron Super 49B v1.5 (Reasoning)	–	80.9%	–	–
51	MiniMax M2	–	80.9%	–	–
52	Kimi K2	none	80.8%	–	–
53	GLM 4.6 V	–	80.6%	–	–
54	INTELLECT-3 (Based on GLM-4.5-Air)	–	80.3%	–	–
55	Mistral Large 3	none	80.3%	–	–
56	GLM 4.7	none	79.9%	–	–
57	Ring-1T	–	79.8%	–	–
58	Doubao Seed Code	–	79.8%	–	–
59	Nova Premier	none	79.6%	–	–
60	Solar Pro 2 (Reasoning)	high	79.6%	–	–
61	Qwen 3 Coder	none	79.4%	–	–
62	K-EXAONE (Reasoning)	–	78.9%	–	–
63	Nova 2.0 Pro Preview (Non-reasoning)	none	78.6%	–	–
64	HyperCLOVA X SEED Think (32B)	–	78.6%	–	–
65	K2-V2 (High)	high	78.6%	–	–
66	Nova 2.0 Lite (Non-reasoning)	none	78.2%	–	–
67	Seed-OSS-36B-Instruct	–	77.9%	–	–
68	Devstral 2	none	77.5%	–	–
69	K2-V2 (Medium)	medium	76.7%	–	–
70	Llama 3.3 70b	none	76.4%	–	–
71	Gemma 3 27B	none	76.4%	–	–
72	Magistral Small 1.2	–	76.3%	–	–
73	Solar Pro 2 (Non-reasoning)	none	75.3%	–	–
74	gpt-oss-20b	high	75.2%	–	–
75	Mi:dm K 2.5 Pro	–	74.2%	–	–
76	Llama 4 Scout	none	74.1%	–	–
77	EXAONE 4.0 32B (Reasoning)	–	73.5%	–	–
78	Nova 2.0 Omni (Non-reasoning)	none	72.7%	–	–
79	Mistral Small 3.2	none	72.1%	–	–
80	Qwen 3 30B-A3B	none	72.0%	–	–
81	Olmo 3.1 32B Think	–	71.5%	–	–
82	K2-V2 (Low)	low	71.4%	–	–
83	K-EXAONE (Non-reasoning)	none	71.0%	–	–
84	Gemma 3 12B	none	70.9%	–	–
85	Apriel-v1.5-15B-Thinker	–	70.4%	–	–
86	Devstral Small 2	none	69.2%	–	–
87	Ministral 3 14B	none	68.3%	–	–
88	Qwen 3 4B	none	67.7%	–	–
89	Ministral 3 8B	none	66.3%	–	–
90	Haiku 3	none	63.3%	–	–
91	NVIDIA Nemotron 3 Nano 30B A3B (Non-reasoning)	none	61.2%	–	–
92	Ministral 3 3B	none	56.7%	–	–
93	Tiny Aya Global	none	55.1%	–	–
94	Gemma 3 4B	none	54.4%	–	–
95	Qwen 3 1.7B	none	48.8%	–	–
96	Gemma 3 1B	none	36.0%	–	–
97	Llama 3.1 8b	none	35.1%	–	–
98	LFM2 2.6B	none	34.3%	–	–