Model rankings

#	Model	Source	Range	LiveBench	Bench
1	Kimi K3.1 ProMoonshotai	Measured	86.8-93.7	unavailable	90.0
2	DeepSeek V4.1 Flash 1017DeepSeek	Measured	83.9-89.8	unavailable	86.7
3	Kimi K3Moonshotai	Measured	83.2-88.4	79.2	85.7
4	GLM 5.3Z.AI	Measured	81.0-86.6	76.1	83.7
5	DeepSeek V4.1 FlashDeepSeek	Measured	80.7-85.7	81.1	83.1
6	DeepSeek V4 Pro 0813DeepSeek	Measured	80.0-85.0	77.4	82.4
7	Qwen3.8 2.4TQwen	Via LiveBench	73.9-88.0	78.5	80.9
8	GLM 5.3 FlashZ.AI	Measured	77.3-82.0	71.6	79.5
9	DeepSeek V4 Flash Vision ExpDeepSeek	Via LiveBench	72.2-86.4	76.8	79.3
10	GLM 5.2Z.AI	Measured	76.6-81.2	73.2	78.8
11	Qwen3.8 Flash NextQwen	Via LiveBench	71.7-85.9	76.2	78.8
12	Qwen3.8 27BQwen	Measured	75.9-81.1	75.3	78.4
13	Kimi K2.6Moonshotai	Measured	74.7-79.2	70.5	76.9
14	Kimi K2.7 CodeMoonshotai	Measured	73.9-79.0	68.4	76.4
15	GLM 5.1Z.AI	Measured	73.4-78.5	70.2	75.9
16	DeepSeek V4 ProDeepSeek	Measured	73.0-77.4	71.6	75.2
17	Kimi K2.5Moonshotai	Measured	71.6-75.4	69.1	73.4
18	Qwen3.5 397B A17BQwen3.5-397b-a17b	Measured	69.7-75.2	unavailable	72.4
19	MiniMax M3MiniMaxAI	Measured	69.5-75.0	67.3	72.2
20	gemma 4 31BGoogle	Via LMArena	66.4-77.0	unavailable	71.7
21	DeepSeek V4 FlashDeepSeek	Measured	67.0-75.2	74.2	71.1
22	Qwen3.6 27BQwen	Measured	68.3-73.6	64.0	70.9
23	GLM 5Z.AI	Measured	68.6-72.5	68.9	70.6
24	DeepSeek V3.2DeepSeek	Measured	68.5-72.2	51.8	70.3
25	Kimi K2 ThinkingMoonshotai	Measured	67.1-71.7	61.6	69.4
26	Qwen3 VL 235B A22B InstructQwen	Via LMArena	63.1-73.7	unavailable	68.4
27	GLM 4.5Z.AI	Via LiveBench	61.0-75.2	65.0	68.1
28	DeepSeek V3.2 ExpDeepSeek	Measured	65.1-70.8	49.9	68.0
29	Qwen3.5 122B A10BQwen	Via LMArena	62.7-73.3	unavailable	68.0
30	Qwen3.5 122B A10BQwen	Via LMArena	62.7-73.3	unavailable	68.0
31	DeepSeek V3.1 TerminusDeepSeek	Via LiveBench	60.7-74.9	64.7	67.8
32	gemma 4 31BGoogle	Measured	64.7-71.0	61.6	67.8
33	Qwen3.6 35B A3BQwen	Measured	65.2-69.9	unavailable	67.6
34	Qwen3.5 35B A3BQwen	Measured	64.3-70.7	unavailable	67.5
35	Mistral Large 4Mistral	Via LMArena	62.3-72.8	unavailable	67.5
36	Qwen3.5 27BQwen	Via LMArena	61.5-72.0	unavailable	66.8
37	Qwen3.5 27BQwen	Via LMArena	61.5-72.0	unavailable	66.8
38	Loes LargeQwen	Via LMArena	61.5-72.0	unavailable	66.8
39	GLM 4.7Z.AI	Measured	64.0-69.0	58.1	66.6
40	gemma 4 26B A4BGoogle	Measured	61.4-67.7	unavailable	64.6
41	Kimi K2 InstructMoonshotai	Measured	60.5-67.5	48.1	64.1
42	GLM 4.5 AirZ.AI	Via LiveBench	56.7-70.9	60.5	63.8
43	DeepSeek V3.1DeepSeek	Measured	59.6-67.5	unavailable	63.6
44	DeepSeek R1 0528DeepSeek	Measured	61.2-65.6	69.4	63.5
45	MiniMax M2.5MiniMaxAI	Via LiveBench	56.3-70.5	60.2	63.4
46	GPT-OSS 120BGpt-oss-120b	Measured	60.4-65.0	46.1	62.8
47	Qwen3.5 9BQwen	Measured	58.9-64.4	unavailable	61.7
48	DeepSeek R1DeepSeek	Measured	58.4-63.1	72.7	60.8
49	Qwen3 235B A22BQwen	Measured	57.5-63.4	63.4	60.5
50	Phi 4 reasoning plusMicrosoft	Via LiveBench	53.0-67.1	56.6	60.1

Sources: Epoch AI and LMArena (CC BY 4.0), LiveBench. Refreshed every morning. epoch.ai ↗

#	Model	Size	MMLU	GPQA	MATH	Code	Index
1	DeepSeek R1 0528DeepSeek	unavailable	91.4	81.0	97.0	92.0	90.4
2	GPT-OSS 120BGpt-oss-120b	unavailable	unavailable	80.1	97.9	unavailable	89.0
3	DeepSeek R1DeepSeek	unavailable	90.8	71.5	97.3	90.0	87.4
4	GLM 4.6Z.AI	unavailable	85.5	82.0	93.5	unavailable	87.0
5	Qwen2.5 32BQwen	32B	83.3	unavailable	83.1	88.4	84.9
6	GLM 4.5Z.AI	unavailable	84.6	79.1	91.0	unavailable	84.9
7	gpt oss 20bOpenAI	21B	unavailable	71.5	96.0	unavailable	83.8
8	Qwen3 235B A22BQwen	235B	87.8	71.1	92.2	unavailable	83.7
9	DeepSeek V3.1DeepSeek	unavailable	90.0	67.0	92.0	85.0	83.5
10	DeepSeek V3 (685B MoE)DeepSeek	685B	89.5	62.7	91.6	84.0	81.9
11	GLM 4.5 AirZ.AI	110B	81.0	75.0	89.0	unavailable	81.7
12	Qwen 2.5 14BQwen	14B	79.7	unavailable	80.0	83.5	81.1
13	DeepSeek V3DeepSeek	unavailable	88.5	59.1	90.2	82.6	80.1
14	DeepSeek R1 Distill Llama 70BDeepSeek	70B	unavailable	65.2	94.5	unavailable	79.9
15	Qwen3 32BQwen	32B	82.4	68.4	88.5	unavailable	79.8
16	DeepSeek R1 Distill 32BDeepSeek	32B	unavailable	62.1	94.3	unavailable	78.2
17	Qwen3 14BQwen	14B	80.7	65.8	87.7	unavailable	78.1
18	Qwen3 30B A3BQwen	30B	79.5	65.8	86.6	unavailable	77.3
19	DeepSeek R1 Distill 14BDeepSeek	14B	unavailable	59.1	93.9	unavailable	76.5
20	Qwen 2.5 72BQwen	72B	86.1	49.0	83.1	86.6	76.2
21	Phi 4 (14B)Microsoft	14B	84.8	56.1	80.4	82.6	76.0
22	Llama 3.3 70BMeta	70B	86.0	50.5	77.0	88.4	75.5
23	Llama 3.1 405B InstructMeta	405B	87.3	50.7	73.8	89.0	75.2
24	Qwen 2.5 Coder 32BQwen	32B	unavailable	unavailable	57.2	92.7	75.0
25	Qwen3 8BQwen	8B	76.9	62.0	85.0	unavailable	74.6
26	Qwen 2.5 Coder 7BQwen	7B	unavailable	unavailable	57.2	88.4	72.8
27	DeepSeek R1 Distill Qwen 7BDeepSeek	7.6B	unavailable	49.1	92.8	unavailable	71.0
28	Llama 3.1 70BMeta	70B	83.6	41.7	68.0	80.5	68.5
29	Qwen2.5 7BQwen	7B	74.2	36.4	75.5	84.8	67.7
30	Gemma 3 27BGoogle	27B	unavailable	42.4	89.0	unavailable	65.7
31	Phi 4 Mini (3.8B)Microsoft	3.8B	67.3	unavailable	64.0	unavailable	65.7
32	gemma 3 12bGoogle	12B	unavailable	34.9	83.8	unavailable	59.3
33	gemma 2 27b itGoogle	27B	75.2	unavailable	42.3	unavailable	58.8
34	Phi 3.5 mini instructMicrosoft	3.8B	69.0	unavailable	48.5	unavailable	58.8
35	Llama 3.1 8BMeta	8B	69.4	30.4	51.9	72.6	56.1
36	Llama 3.2 3BMeta	3B	63.4	unavailable	48.0	unavailable	55.7
37	gemma 2 9b itGoogle	9.2B	71.3	unavailable	36.6	unavailable	54.0
38	Gemma 3 4BGoogle	4B	unavailable	30.8	75.6	unavailable	53.2
39	Mistral Nemo 12BEUMistral	12B	68.0	unavailable	38.0	unavailable	53.0
40	Llama 3.2 1BMeta	1B	49.3	unavailable	30.6	unavailable	40.0
41	Mistral 7B Instruct v0.2EUMistral	7.2B	60.1	unavailable	11.2	30.5	33.9

Benchmarks are publicly reported figures from model authors and open leaderboards (last updated 2026-06). Methodology varies per model, so the index is an ordering aid, not a controlled benchmark.

#	Model	Size	Response
1	codegemma 2bGoogle	2.5B	66 ms
2	gemma 2b itGoogle	2.5B	67 ms
3	Llama 3.2 3B InstructMeta	3.2B	76 ms
4	Phi 3 mini 128k instructMicrosoft	3.8B	81 ms
5	Apertus v1.1 1.5B InstructEUSwiss-ai	1.5B	81 ms
6	Flor 6.3B Instruct 4096EUBSC-LT	6.2B	81 ms
7	Phi 3.5 mini instructMicrosoft	3.8B	82 ms
8	granite vision 3.1 2b previewIbm-granite	3B	83 ms
9	Llama 3.2 1B InstructMeta	1.2B	87 ms
10	Meta Llama 3 8B InstructMeta	8B	89 ms
11	Qwen3 VL 2B InstructQwen	2.1B	93 ms
12	Phi 3.5 vision instructMicrosoft	4.1B	93 ms
13	Flor 6.3B InstructEUBSC-LT	6.2B	94 ms
14	Qwen3.5 0.8BQwen	0.9B	96 ms
15	Qwen3.5 2BQwen	2B	96 ms
16	Qwen2.5 14B InstructQwen	15B	98 ms
17	phi 4Microsoft	15B	98 ms
18	Qwen2.5 VL 7B Surg CholecT50NVIDIA	8.3B	101 ms
19	salamandra 7b instruct tools 16kEUBSC-LT	7.8B	102 ms
20	Phi 4 mini instructMicrosoft	3.8B	103 ms
21	salamandra 7b instruct 2606EUBSC-LT	7.8B	103 ms
22	X Reasoner 7BMicrosoft	8.3B	106 ms
23	EuroLLM 22B Instruct PreviewEUUtter-project	23B	109 ms
24	Llama Poro 2 8B SFTEULumiOpen	8B	109 ms
25	EuroVLM 1.7B PreviewEUUtter-project	2.1B	110 ms

Response times were measured at our last verification of each model and vary with load and cold starts.