Best LLMs for science: open and proprietary

Open and proprietary models together on one 0–100 scale (scientific knowledge). Switch to local models to see only what you can download and run.

40 benchmarks4 sourcesUpdated 3 Oct 2026Reference scale 2026-Q4How the score is built

Local models Overall

General Coding Agents Math Science Reasoning

Models ranked by science llmrun Score
#	Model	llmrun Score	Coding	Agents	Math	Science	Reasoning	Benchmarks	VRAM
1	Claude Opus 5.54 benchmarks		–	–	–	59	74	4	–
2	Claude Opus 5.5 Max17 benchmarks		78	–	96	58	82	17	–
3	Claude Sonnet 5.5 Max13 benchmarks		75	–	93	58	81	13	–
4	Claude Fable 5.1 Max15 benchmarks		79	–	95	57	75	15	–
5	Claude Sonnet 5.53 benchmarks	—	75	–	–	57	–	3	–
6	GPT 5.6 Sol Max21 benchmarks		78	70	91	57	77	21	–
7	GPT 6 Astra Max19 benchmarks		83	72	96	57	86	19	–
8	Claude Fable 5.19 benchmarks		78	–	–	56	74	9	–
9	GPT 6 Sol Max18 benchmarks		79	–	94	56	75	18	–
10	GPT 6.1 Sol Max13 benchmarks		64	–	96	56	83	13	–
11	GPT 5.4 Pro (Mar 05, 2026)8 benchmarks		–	–	83	56	74	8	–
12	GPT 5.6 Sol9 benchmarks		75	–	–	55	72	9	–
13	Claude Opus 5 Max22 benchmarks		78	72	91	55	75	22	–
14	GPT 5.6 Terra Max20 benchmarks		73	67	87	55	71	20	–
15	GPT 6 Sol4 benchmarks		–	–	–	54	68	4	–
16	Claude Opus 59 benchmarks		78	–	–	54	74	9	–
17	Claude Fable 5 Max20 benchmarks		78	62	94	54	74	20	–
18	GPT 6.1 Sol3 benchmarks	—	73	–	–	54	–	3	–
19	Muse Spark 1.3 Max6 benchmarks		–	–	78	54	61	6	–
20	GPT 5.59 benchmarks		74	–	–	53	72	9	–
21	GPT 5.6 Terra10 benchmarks		70	–	–	53	65	10	–
22	Kimi K3open1674 GB · 26 benchmarks		73	61	81	52	63	26	1674 GB
23	Step 5 Preview2 benchmarks	—	–	–	–	52	–	2	–
24	GPT 5.4 (Mar 05, 2026)20 benchmarks		69	57	78	51	67	20	–
25	GLM 5.3open456 GB · 21 benchmarks		67	68	72	49	58	21	456 GB
26	Muse Spark 1.27 benchmarks		59	–	–	49	62	7	–
27	Grok 4.79 benchmarks		–	–	63	49	63	9	–
28	Muse Spark 1.19 benchmarks		66	59	69	48	–	9	–
29	GPT 5.6 Luna Max20 benchmarks		69	63	82	48	62	20	–
30	Claude Opus 4.8 Max19 benchmarks		64	55	82	48	66	19	–
31	GPT 5.6 Luna8 benchmarks		–	–	–	48	58	8	–
32	Gemini 3.7 Flash15 benchmarks		56	–	74	48	70	15	–
33	GPT 6 Luna Max18 benchmarks		69	–	82	47	60	18	–
34	Muse Spark6 benchmarks		–	–	58	47	–	6	–
35	GLM 5.2open456 GB · 24 benchmarks		58	60	68	47	52	24	456 GB
36	Grok 4.516 benchmarks		65	–	66	47	60	16	–
37	GPT 6 Luna5 benchmarks		67	–	–	47	55	5	–
38	DeepSeek V4 Pro 0813open991 GB · 21 benchmarks		63	–	71	46	65	21	991 GB
39	Gemini 3 Pro Preview17 benchmarks		62	55	60	46	54	17	–
40	Gemini 3.6 Flash16 benchmarks		52	–	66	45	63	16	–
41	Gemini 3.8 Flash6 benchmarks		69	–	–	45	65	6	–
42	DeepSeek V4.1 Flashopen458 GB · 10 benchmarks		54	–	76	45	57	10	458 GB
43	Claude Opus 4.6 Max12 benchmarks		–	–	71	45	55	12	–
44	DeepSeek V4 Flash 0731open183 GB · 21 benchmarks		59	–	71	45	62	21	183 GB
45	GLM 5.3 Flashopen195 GB · 17 benchmarks		52	63	62	44	37	17	195 GB
46	DeepSeek V4 Proopen960 GB · 22 benchmarks		49	–	57	44	49	22	960 GB
47	Qwen3.7 Max17 benchmarks		59	–	68	43	58	17	–
48	Kimi K2.6open620 GB · 18 benchmarks		57	–	63	43	51	18	620 GB
49	Claude Opus 4.7 Max15 benchmarks		67	–	73	42	64	15	–
50	Grok Build 0.15 benchmarks		50	–	–	41	–	5	–
51	Inkling Smallopen160 GB · 12 benchmarks		–	–	57	41	54	12	160 GB
52	MiniMax M3open257 GB · 19 benchmarks		52	40	46	41	38	19	257 GB
53	Grok 4.311 benchmarks		–	–	57	41	52	11	–
54	GPT 5 (Aug 07, 2025)20 benchmarks		58	–	62	40	48	20	–
55	Kimi K2.7 Codeopen620 GB · 19 benchmarks		53	46	61	40	50	19	620 GB
56	GPT 5.4 Mini (Mar 17, 2026)12 benchmarks		–	–	60	40	47	12	–
57	Kimi K2.5open620 GB · 20 benchmarks		49	37	–	40	45	20	620 GB
58	GLM 5.1open457 GB · 13 benchmarks		53	54	57	40	46	13	457 GB
59	Qwen3.7 Plus9 benchmarks		42	–	–	40	46	9	–
60	Inklingopen572 GB · 21 benchmarks		43	–	52	40	53	21	572 GB
61	GPT 5.4 Nano (Mar 17, 2026)6 benchmarks		–	–	–	39	42	6	–
62	MiMo V2.5 Proopen615 GB · 6 benchmarks		54	–	–	38	42	6	615 GB
63	Claude Sonnet 5 Max13 benchmarks		–	–	77	38	57	13	–
64	Qwen3.6 Plus15 benchmarks		47	–	55	38	41	15	–
65	Gemini 2.5 Pro Preview (Jun 05)6 benchmarks		–	–	–	38	–	6	–
66	Gemini 2.5 Pro13 benchmarks		–	29	49	37	43	13	–
67	Qwen3.8 27Bopen17.4 GB · 9 benchmarks		51	–	57	37	50	9	17.4 GB
68	NVIDIA Nemotron 3 Ultra 550B A55B BF16open370 GB · 14 benchmarks		43	–	52	37	46	14	370 GB
69	Gemini 2.5 Pro Exp (Mar 25)6 benchmarks		–	–	–	37	–	6	–
70	DeepSeek V4 Flashopen175 GB · 10 benchmarks		46	–	–	36	43	10	175 GB
71	GLM 4.7open216 GB · 13 benchmarks		41	32	51	36	35	13	216 GB
72	GPT 5.5 Instant7 benchmarks		–	–	45	36	–	7	–
73	Qwen3.6 27Bopen17.4 GB · 12 benchmarks		42	–	53	36	37	12	17.4 GB
74	O3 (Apr 16, 2025)19 benchmarks		52	–	52	35	46	19	–
75	Qwen3.6 35B A3Bopen22.0 GB · 12 benchmarks		44	–	49	35	38	12	22.0 GB
76	MiniMax M2.7open138 GB · 6 benchmarks		41	–	–	35	–	6	138 GB
77	Claude Sonnet 4.6 Max11 benchmarks		–	–	70	34	59	11	–
78	Qwen3 235B A22B Thinking 2507open142 GB · 13 benchmarks		42	–	–	34	37	13	142 GB
79	O4 Mini (Apr 16, 2025)19 benchmarks		49	–	52	34	42	19	–
80	Qwen3.5 35B A3Bopen22.0 GB · 8 benchmarks		–	–	–	34	37	8	22.0 GB
81	MiMo V2.5open187 GB · 4 benchmarks		43	–	–	33	–	4	187 GB
82	Ring 2.6 1Topen677 GB · 3 benchmarks	—	40	–	–	33	–	3	677 GB
83	Claude Sonnet 4.5 (Sep 29, 2025)11 benchmarks		–	42	–	32	46	11	–
84	O3 Mini (Jan 31, 2025)18 benchmarks		42	–	47	32	31	18	–
85	Gemma 4 31B ITopen20.4 GB · 11 benchmarks		47	–	–	32	40	11	20.4 GB
86	Nova 2.0 Pro Preview2 benchmarks	—	–	–	–	32	–	2	–
87	Qwen3.5 9Bopen6.4 GB · 9 benchmarks		–	–	–	32	30	9	6.4 GB
88	GPT OSS 120Bopen70.5 GB · 20 benchmarks		37	14	–	31	30	20	70.5 GB
89	DeepSeek V3.1 Terminusopen415 GB · 5 benchmarks		47	–	–	31	39	5	415 GB
90	Step 3.7 Flashopen133 GB · 3 benchmarks	—	46	–	–	30	–	3	133 GB
91	Cogito 671B V2.1open407 GB · 2 benchmarks	—	–	–	–	30	–	2	407 GB
92	Gemma 4 26B A4B ITopen16.1 GB · 9 benchmarks		43	–	–	30	33	9	16.1 GB
93	Mistral Medium 26049 benchmarks		41	–	–	30	33	9	–
94	DeepSeek V3.2 Expopen415 GB · 10 benchmarks		46	–	–	30	44	10	415 GB
95	Magistral Medium 25092 benchmarks	—	–	–	–	29	–	2	–
96	Mercury 24 benchmarks		45	–	–	29	–	4	–
97	GLM 4.6open215 GB · 6 benchmarks		36	31	–	29	–	6	215 GB
98	Qwen3 30B A3B Thinking 2507open18.7 GB · 8 benchmarks		–	–	–	28	27	8	18.7 GB
99	Command A Plus 05 2026 BF16open132 GB · 2 benchmarks	—	–	–	–	28	–	2	132 GB
100	DeepSeek R1open415 GB · 14 benchmarks		39	–	40	28	29	14	415 GB
101	Gemini 2.5 Pro Preview (May 06)5 benchmarks		–	–	–	27	–	5	–
102	GPT 4.5 Preview (Feb 27, 2025)11 benchmarks		36	–	33	27	28	11	–
103	NVIDIA Nemotron 3 Super 120B A12B BF16open74.7 GB · 4 benchmarks		36	–	–	27	–	4	74.7 GB
104	DeepSeek v3 0324open415 GB · 13 benchmarks		39	–	33	27	24	13	415 GB
105	GPT 4.1 Mini (Apr 14, 2025)16 benchmarks		32	–	36	27	26	16	–
106	Trinity Large Thinkingopen263 GB · 3 benchmarks	—	32	–	–	26	–	3	263 GB
107	GPT 4.1 (Apr 14, 2025)17 benchmarks		39	–	34	26	26	17	–
108	GPT 5.1 2025 11.137 benchmarks		–	–	–	26	27	7	–
109	Mistral Large 3 675B Instruct 2512open446 GB · 5 benchmarks		33	–	–	26	25	5	446 GB
110	Qwen3 32Bopen20.3 GB · 9 benchmarks		33	–	–	26	26	9	20.3 GB
111	Llama 4 Maverick 17B 128E Instructopen265 GB · 17 benchmarks		25	–	29	26	23	17	265 GB
112	Qwen3.5 122B A10Bopen82.6 GB · 6 benchmarks		–	–	–	26	43	6	82.6 GB
113	Claude Opus 4 (May 14, 2025)7 benchmarks		–	–	–	26	46	7	–
114	Qwen3 14Bopen9.5 GB · 8 benchmarks		–	–	–	25	24	8	9.5 GB
115	Mistral Medium 25086 benchmarks		33	–	–	24	26	6	–
116	GPT OSS 20Bopen12.9 GB · 11 benchmarks		40	–	–	24	26	11	12.9 GB
117	Qwen3 Coder Nextopen48.2 GB · 3 benchmarks	—	35	–	–	23	–	3	48.2 GB
118	Mistral Medium 25055 benchmarks		–	–	33	23	–	5	–
119	DeepSeek v3open415 GB · 10 benchmarks		37	–	26	22	21	10	415 GB
120	Gemini 2.0 Flash Thinking Exp (Jan 21)6 benchmarks		–	–	–	22	–	6	–

The range after each score is a 90% interval: given the boards a model has results on, its true score is very likely inside it. Short ranges mean many agreeing results; long ranges mean few or conflicting ones.

Why some models are missing. A model gets a score only with results on at least 4 benchmarks across at least 2 categories, one of them reasoning or coding. Skill scores need at least 2 benchmarks in that skill; otherwise they show as –.

VRAM is llmrun's estimate at Q4_K_M (or the smallest quantization we track) for the weights plus a working context. Proprietary models can't run locally, so the VRAM filter hides them.

How the llmrun Score is built · llmrun does not run these benchmarks; scores are aggregated from public sources.