Instruction following

AA-IFBench leaderboard

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap

As of 2026-10-07, the highest measured score on AA-IFBench is 82.9% by MiniMax M3. 38 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#	Model	Score	Source
1	MiniMax M3	82.9%	measured
2	Nemotron 3 Ultra	81.4%	measured
3	Grok 4.3	81.3%	measured
4	Qwen3.7 Max	80.5%	measured
5	MAI-Thinking-1	80.3%	estimated ± 8.5 pp, low confidence
6	Qwen3.8 Max	80.3%	estimated ± 8.5 pp, low confidence
7	Inkling-Small	80.3%	estimated ± 8.5 pp, low confidence
8	Qwen3.8-Omni-Flash	80.3%	estimated ± 8.5 pp, medium confidence
9	Qwen3.8-Flash-Next	80.3%	estimated ± 8.5 pp, medium confidence
10	Solar Open 2	80.3%	estimated ± 8.5 pp, medium confidence
11	Inkling	80.3%	estimated ± 8.5 pp, medium confidence
12	Beam	80.3%	estimated ± 8.5 pp, medium confidence
13	Qwen3.8-27B	80.3%	estimated ± 8.5 pp, medium confidence
14	Granite 4.2 8B	80.3%	estimated ± 8.5 pp, medium confidence
15	Agents-A1	79.9%	estimated ± 8.2 pp, medium confidence
16	MiMo-V2.5-Pro	79.9%	measured
17	Agents-A1-4B	79.8%	estimated ± 8.2 pp, medium confidence
18	Granite 4.2 30B	79.1%	estimated ± 8.5 pp, medium confidence
19	Mercury 2.5	78.8%	estimated ± 8.5 pp, medium confidence
20	Muse Glimmer 30B	78.8%	estimated ± 8.5 pp, medium confidence
21	Qwen3.7 Plus	78.0%	measured
22	GPT-5.2-Codex	77.6%	measured
23	Gemini 3.1 Pro	77.1%	measured
24	Qwen 3.6 Max (preview)	76.6%	measured
25	DeepSeek V4 Pro 0813	76.5%	measured
26	Gemini 3.5 Flash	76.3%	measured
27	GLM-5.1	76.3%	measured
28	Kimi K2.6	76.0%	measured
29	GPT-5.4 nano	75.9%	measured
30	GPT-5.5	75.9%	measured
31	Muse Spark	75.9%	measured
32	MiniMax M2.7	75.7%	measured
33	Qwen3.5-122B-A10B	75.7%	measured
34	Gemma 4 31B	75.6%	measured
35	Qwen3.5-27B	75.6%	measured
36	GPT-5.2	75.4%	measured
37	GPT-5.3 Codex	75.4%	measured
38	Qwen3.6 Plus	75.2%	measured
39	A.X K2	75.1%	estimated ± 8.5 pp, medium confidence
40	Command A+	73.9%	measured
41	GPT-5.4	73.9%	measured
42	Gemma 4 12B	73.5%	measured
43	GLM-5.2	73.3%	measured
44	GPT-5.4 mini	73.3%	measured
45	GLM-5-Turbo	73.2%	measured
46	GPT-5 (high)	73.1%	measured
47	GPT-5.1	72.9%	measured
48	dots3-note Preview	72.8%	estimated ± 8.2 pp, medium confidence
49	o3-mini	72.8%	estimated ± 8.2 pp, medium confidence
50	GPT-5.6 Sol	72.7%	measured
51	Qwen3.5-35B-A3B	72.5%	measured
52	Gemma 4 26B A4B	72.4%	measured
53	GLM-5	72.3%	measured
54	Nemotron 3 Super 100B	71.5%	measured
55	o3	71.4%	measured
56	GPT-5.6 Terra	71.2%	measured
57	Solar Pro 3	71.2%	measured
58	Nemotron 3 Nano 30B	71.1%	measured
59	GPT-5 (medium)	70.6%	measured
60	Gemini 3 Pro	70.4%	measured
61	o1	70.3%	measured
62	Kimi K2.5	70.2%	measured
63	Kimi K2.5 (Reasoning)	70.2%	measured
64	GPT-5.1-Codex	70.0%	measured
65	GPT-5.1-Codex-Max	70.0%	measured
66	GPT-OSS 120B	69.0%	measured
67	MiMo-V2-Pro	68.8%	measured
68	Mistral Medium 3.5 128B	68.8%	measured
69	GLM-4.7	67.9%	measured
70	Qwen3.6-27B	67.6%	measured
71	Step 3.7 Flash	67.3%	measured
72	Ling 3.0 Flash	65.4%	estimated ± 8.5 pp, medium confidence
73	GPT-OSS 20B	65.1%	measured
74	K-Exaone	64.7%	measured
75	Qwen3.6-35B-A3B	64.4%	measured
76	Granite 4.2 3B	64.2%	estimated ± 8.5 pp, medium confidence
77	Claude Fable 5	63.5%	measured
78	Nemotron 3 Nano Omni 30B A3B	63.2%	measured
79	Kimi K2.7 Code	63.1%	measured
80	K-EXAONE 2.0	63.0%	estimated ± 8.2 pp, medium confidence
81	Claude Opus 4.8	62.2%	measured
82	GLM-5V-Turbo	61.1%	measured
83	Ling 3.0 Flash FP8	59.7%	estimated ± 8.5 pp, medium confidence
84	Claude Opus 4.7 (Adaptive)	58.6%	measured
85	Nemotron 3.5 Lightning 30B A3B NVFP4	58.4%	estimated ± 8.5 pp, medium confidence
86	Claude Opus 4.5 Thinking	58.0%	measured
87	North Mini Code	57.6%	measured
88	Ling 2.6 Flash	57.4%	measured
89	Ternary Bonsai 2 27B	57.0%	estimated ± 8.2 pp, medium confidence
90	Hy3 Preview	56.8%	estimated ± 8.5 pp, medium confidence
91	LFM2.5-2.6B	56.8%	estimated ± 8.5 pp, medium confidence
92	LLaDA2.2-mini	56.8%	estimated ± 8.5 pp, low confidence
93	Trinity-Large-Preview	56.3%	measured
94	Trinity-Large-Thinking	56.3%	measured
95	LFM2.5-8B-A1B	55.6%	measured
96	Claude 4.1 Opus Thinking	55.4%	measured
97	Gemini 3 Flash	55.1%	measured
98	Grok 4	53.7%	measured
99	MiMo-V2-Omni	53.5%	measured
100	Claude Opus 4.6 (Adaptive)	53.1%	measured
101	Grok 4.1 Fast (Reasoning)	52.7%	measured
102	Qwen3.5 397B	51.6%	measured
103	Qwen3.5 397B (Reasoning)	51.6%	measured
104	Grok 4 Fast (Reasoning)	50.5%	measured
105	DeepSeek V3.2	49.0%	measured
106	Gemini 2.5 Pro	48.7%	measured
107	Mistral Small 4	48.2%	measured
108	Mistral Small 4 (Reasoning)	48.2%	measured
109	Ultravox v0.6 Llama 3.3 70B	47.1%	measured
110	Claude 4 Sonnet	45.4%	measured
111	Claude Opus 4.6	44.6%	measured
112	Gemma 4 E4B	44.2%	measured
113	Qwen3 Max	44.1%	measured
114	Claude Opus 4.7	43.6%	measured
115	Qwen3-Omni-30B-A3B-Thinking	43.4%	measured
116	Claude Opus 4.5	43.0%	measured
117	GPT-4.1	43.0%	measured
118	Llama 4 Maverick	43.0%	measured
119	DeepSeek V3.1 (Reasoning)	41.5%	measured
120	Kimi K2	41.5%	measured
121	Grok Code Fast 1	41.4%	measured
122	Claude Sonnet 4.6	41.2%	measured
123	DeepSeek V3 0324	41.0%	measured
124	MiMo-V2-Flash	39.9%	measured
125	DeepSeek-R1	39.6%	measured
126	Llama 4 Scout	39.5%	measured
127	Mistral Medium 3	39.3%	measured
128	Gemini 2.5 Flash	39.0%	measured
129	Llama 3.1 405B	39.0%	measured
130	GPT-4.1 mini	38.3%	measured
131	MiniCPM5-2B	38.2%	estimated ± 8.2 pp, medium confidence
132	Nemotron Ultra 253B	38.2%	measured
133	Nova Pro	38.1%	measured
134	Gemma 4 E2B	38.0%	measured
135	DeepSeek V3.1	37.8%	measured
136	GLM-4.5-Air	37.6%	measured
137	GLM-4.6	36.7%	measured
138	Grok 4.1 Fast	36.5%	measured
139	Mistral Large 3	36.2%	measured
140	Claude 3 Haiku	36.1%	measured
141	ZAYA1-8B	34.8%	estimated ± 8.2 pp, medium confidence
142	DeepSeek V3	34.8%	measured
143	Sarvam 105B	34.4%	measured
144	GPT-4o	34.3%	measured
145	Solar Pro 2	33.7%	measured
146	Exaone 4.0 32B	33.5%	measured
147	LFM2.5-VL-1.6B-Extract	33.1%	measured
148	GPT-4.1 nano	32.0%	measured
149	Gemma 3 27B	31.8%	measured
150	Mistral Large 2	31.2%	measured
151	Qwen3-Omni-30B-A3B-Instruct	31.2%	measured
152	GPT-4o mini	31.0%	measured
153	LFM2.5-VL-3B	26.5%	estimated ± 8.2 pp, low confidence
154	Sarvam 30B	26.5%	measured
155	Granite-4.0-H-1B	26.2%	measured
156	Exaone 4.0 1.2B	25.3%	measured
157	Kanana-2 3B Instruct	23.7%	estimated ± 8.2 pp, low confidence
158	Phi-4	23.5%	measured
159	Celeris-1	23.4%	estimated ± 8.2 pp, low confidence
160	DeepSeek R1 Distill Qwen 32B	22.9%	measured
161	MiniCPM5-1B	22.6%	estimated ± 8.2 pp, low confidence
162	Kanana-2 1.3B Instruct	17.8%	estimated ± 8.2 pp, low confidence
163	Granite-4.0-H-350M	17.6%	measured
164	Mellum2-12B-A2.5B-Thinking	16.1%	estimated ± 8.2 pp, low confidence
165	Granite-4.0-350M	15.9%	measured
166	Mellum2-12B-A2.5B-Instruct	15.0%	estimated ± 8.2 pp, low confidence
167	LFM2.5-230M	9.9%	estimated ± 8.2 pp, low confidence
168	LFM2.5-VL-450M	0.8%	estimated ± 8.2 pp, low confidence

AA-IFBench leaderboard

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General