llmrun Composite Score · Science

The best open LLMs for science, ranked on graduate-level science and knowledge boards such as GPQA. The score shown is the overall llmrun Composite Score; the highlighted column is the scientific knowledge skill score used for the order.

45 benchmarks4 sourcesUpdated 17 Oct 2026Reference scale 2026-Q4How the Composite Score is built

Science score vs VRAM needed

This chart doesn't show proprietary models.

Overall Open models

General Coding Agents Math Science Reasoning

Models ranked by science llmrun Composite Score
# (open / overall)	Model	llmrun Composite Score	Coding	Agents	Math	Science	Reasoning	Benchmarks	VRAM
1 / 11	MiMo V2.6 Pro RL614 GB · 8 benchmarks		126	132	–	140	–	8	614 GB
2 / 19	Kimi K31670 GB · 30 benchmarks		134	131	135	136	129	30	1670 GB
3 / 23	GLM 5.3454 GB · 26 benchmarks		130	133	129	131	130	26	454 GB
4 / 31	GLM 5.2454 GB · 26 benchmarks		120	128	126	129	120	26	454 GB
5 / 38	DeepSeek V4 Pro 0813991 GB · 21 benchmarks		125	–	128	126	131	21	991 GB
6 / 42	DeepSeek V4 Flash 0731183 GB · 21 benchmarks		121	–	128	124	129	21	183 GB
7 / 44	GLM 5.3 Flash205 GB · 24 benchmarks		117	127	121	124	127	24	205 GB
8 / 45	DeepSeek V4.1 Flash458 GB · 24 benchmarks		122	128	129	123	134	24	458 GB
9 / 47	DeepSeek V4 Pro960 GB · 22 benchmarks		111	–	118	122	117	22	960 GB
10 / 49	MiMo V2.6 Flash RL187 GB · 4 benchmarks		124	–	–	122	–	4	187 GB
11 / 50	Kimi K2.6618 GB · 20 benchmarks		119	122	122	122	118	20	618 GB
12 / 54	Inkling Small163 GB · 11 benchmarks		–	–	117	118	123	11	163 GB
13 / 55	MiniMax M3258 GB · 22 benchmarks		115	112	106	118	108	22	258 GB
14 / 57	Kimi K2.7 Code618 GB · 19 benchmarks		115	119	121	118	115	19	618 GB
15 / 58	GLM 5.1454 GB · 14 benchmarks		115	123	117	117	116	14	454 GB
16 / 60	Kimi K2.5618 GB · 19 benchmarks		110	111	–	117	115	19	618 GB
17 / 62	Inkling575 GB · 25 benchmarks		110	111	112	116	121	25	575 GB
18 / 65	MiMo V2.5 Pro616 GB · 6 benchmarks		116	–	–	115	111	6	616 GB
19 / 70	MiniMax M2.7140 GB · 7 benchmarks		103	112	–	112	–	7	140 GB
20 / 72	Qwen3.8 27B17.9 GB · 15 benchmarks		120	123	118	112	124	15	17.9 GB
21 / 73	NVIDIA Nemotron 3 Ultra 550B A55B BF16378 GB · 17 benchmarks		104	102	112	112	116	17	378 GB
22 / 75	DeepSeek V4 Flash175 GB · 10 benchmarks		107	–	–	111	112	10	175 GB
23 / 76	GLM 4.7218 GB · 13 benchmarks		101	106	111	111	111	13	218 GB
24 / 77	Qwen3.6 27B17.9 GB · 11 benchmarks		102	–	114	111	105	11	17.9 GB
25 / 78	Qwen3.6 35B A3B22.0 GB · 11 benchmarks		105	–	109	110	104	11	22.0 GB
26 / 82	Ring 2.6 1T687 GB · 3 benchmarks	—	102	–	–	109	–	3	687 GB
27 / 84	Muse Glimmer 30B19.6 GB · 2 benchmarks	—	–	–	–	108	–	2	19.6 GB
28 / 86	Qwen3.5 35B A3B22.0 GB · 7 benchmarks		–	–	–	108	107	7	22.0 GB
29 / 88	Qwen3 235B A22B Thinking 2507143 GB · 12 benchmarks		102	–	–	108	106	12	143 GB
30 / 90	MiMo V2.5188 GB · 4 benchmarks		103	–	–	107	–	4	188 GB
31 / 92	Gemma 4 31B IT21.7 GB · 13 benchmarks		108	–	–	105	110	13	21.7 GB
32 / 94	Qwen3.5 9B6.6 GB · 8 benchmarks		–	–	–	105	97	8	6.6 GB
33 / 101	GPT OSS 120B70.8 GB · 19 benchmarks		98	82	–	103	96	19	70.8 GB
34 / 102	Step 3.7 Flash139 GB · 3 benchmarks	—	107	–	–	102	–	3	139 GB
35 / 103	Cogito 671B V2.1404 GB · 2 benchmarks	—	–	–	–	102	–	2	404 GB
36 / 104	Gemma 4 26B A4B IT16.5 GB · 8 benchmarks		104	–	–	102	102	8	16.5 GB
37 / 106	DeepSeek V3.2 Exp413 GB · 10 benchmarks		107	–	–	101	113	10	413 GB
38 / 110	GLM 4.6217 GB · 6 benchmarks		96	104	–	100	–	6	217 GB
39 / 111	Command A Plus 05 2026 BF16132 GB · 2 benchmarks	—	–	–	–	100	–	2	132 GB
40 / 112	DeepSeek V3.1 Terminus412 GB · 5 benchmarks		107	–	–	99	108	5	412 GB
41 / 114	DeepSeek R1412 GB · 13 benchmarks		100	–	–	99	96	13	412 GB
42 / 115	Qwen3 30B A3B Thinking 250719.4 GB · 7 benchmarks		–	–	–	99	94	7	19.4 GB
43 / 116	DeepSeek v3 0324412 GB · 12 benchmarks		100	–	–	98	88	12	412 GB
44 / 119	NVIDIA Nemotron 3 Super 120B A12B BF1674.7 GB · 4 benchmarks		96	–	–	98	–	4	74.7 GB
45 / 121	Mistral Large 3 675B Instruct 2512454 GB · 5 benchmarks		93	–	–	97	89	5	454 GB
46 / 122	Trinity Large Thinking271 GB · 3 benchmarks	—	92	–	–	97	–	3	271 GB
47 / 125	Qwen3 32B22.6 GB · 8 benchmarks		93	–	–	96	92	8	22.6 GB
48 / 126	Qwen3.5 122B A10B75.5 GB · 6 benchmarks		–	–	–	96	112	6	75.5 GB
49 / 127	Llama 4 Maverick 17B 128E Instruct273 GB · 16 benchmarks		81	–	–	96	85	16	273 GB
50 / 128	Qwen3 14B11.9 GB · 7 benchmarks		–	–	–	94	88	7	11.9 GB
51 / 129	GPT OSS 20B13.1 GB · 10 benchmarks		101	–	–	92	91	10	13.1 GB
52 / 131	Qwen3 Coder Next48.3 GB · 3 benchmarks	—	95	–	–	91	–	3	48.3 GB
53 / 133	DeepSeek v3412 GB · 8 benchmarks		97	–	–	90	81	8	412 GB
54 / 137	Qwen3 8B7.6 GB · 7 benchmarks		–	–	–	88	81	7	7.6 GB
55 / 138	Devstral Small 2 24B Instruct 251218.1 GB · 3 benchmarks	—	–	–	–	86	–	3	18.1 GB
56 / 139	Llama 4 Scout 17B 16E Instruct77.0 GB · 11 benchmarks		–	–	–	83	80	11	77.0 GB
57 / 140	Magistral Small 250918.1 GB · 5 benchmarks		–	–	–	83	–	5	18.1 GB
58 / 141	Mistral Small 3.2 24B Instruct 250618.1 GB · 5 benchmarks		–	–	–	82	–	5	18.1 GB
59 / 142	MiMo v2 Flash187 GB · 3 benchmarks	—	100	–	–	82	–	3	187 GB
60 / 144	Granite 4.1 30B21.9 GB · 2 benchmarks	—	–	–	–	82	–	2	21.9 GB
61 / 147	Mistral Small 3.1 24B Instruct 250318.1 GB · 5 benchmarks		–	–	–	81	–	5	18.1 GB
62 / 148	Llama 3.3 70B Instruct51.3 GB · 12 benchmarks		76	–	–	81	82	12	51.3 GB
63 / 149	Gemma 3 27B IT21.7 GB · 9 benchmarks		70	–	–	81	74	9	21.7 GB
64 / 151	Gemma 3 12B IT10.9 GB · 7 benchmarks		–	–	–	71	66	7	10.9 GB
65 / 152	Phi 4 Mini Instruct4.8 GB · 3 benchmarks	—	–	–	–	61	–	3	4.8 GB
66 / 153	Llama 3.1 8B Instruct7.9 GB · 9 benchmarks		56	–	–	56	69	9	7.9 GB

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 receives a score only if it has results on at least 4 benchmarks across at least 2 categories, with one being reasoning or coding, or on 3 benchmarks that narrow its 90% range to at most 15.6 points. The results must come from at least two different benchmarks: categories of one benchmark, such as LiveBench, count once. Skill scores require at least 2 benchmarks for that skill; otherwise they display as –.

VRAM is llmrun's estimate at Q4_K_M (or the smallest quantization we track) for the weights plus room for a 16K-token context. Longer contexts need more memory. Proprietary models can't run locally, so the VRAM filter hides them.

How the llmrun Composite Score is built · llmrun doesn't run these benchmarks; scores come from public sources.

Dataset. Grab the scores of every open model as a dataset on Hugging Face, including a timestamped snapshot for each refresh.