llmrun Composite Score · Reasoning

The best open LLMs for reasoning, ranked on hard multi-step reasoning boards such as ARC-AGI and HLE. The score shown is the overall llmrun Composite Score; the highlighted column is the reasoning skill score used for the order.

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

Reasoning score vs VRAM needed

This chart doesn't show proprietary models.

Open models Overall

General Coding Agents Math Science Reasoning

Models ranked by reasoning llmrun Composite Score
# (open / overall)	Model	llmrun Composite Score	Coding	Agents	Math	Science	Reasoning	Benchmarks	VRAM
1 / 28	DeepSeek V4.1 Flash458 GB · 20 benchmarks		119	–	129	123	133	20	458 GB
2 / 37	DeepSeek V4 Flash 0731183 GB · 21 benchmarks		121	–	128	124	129	21	183 GB
3 / 43	GLM 5.3 Flash205 GB · 21 benchmarks		113	131	121	124	127	21	205 GB
4 / 48	GLM 5.3454 GB · 23 benchmarks		131	138	129	131	126	23	454 GB
5 / 53	Qwen3.8 27B17.9 GB · 12 benchmarks		112	–	118	112	124	12	17.9 GB
6 / 58	Inkling Small163 GB · 12 benchmarks		–	–	117	118	122	12	163 GB
7 / 61	Inkling575 GB · 22 benchmarks		104	109	112	116	121	22	575 GB
8 / 64	GLM 5.2454 GB · 26 benchmarks		120	128	126	129	120	26	454 GB
9 / 65	Kimi K2.6618 GB · 20 benchmarks		119	122	122	122	119	20	618 GB
10 / 69	Kimi K2.7 Code618 GB · 19 benchmarks		115	119	121	118	118	19	618 GB
11 / 76	NVIDIA Nemotron 3 Ultra 550B A55B BF16378 GB · 15 benchmarks		103	–	112	112	115	15	378 GB
12 / 78	GLM 5.1454 GB · 14 benchmarks		115	123	117	117	115	14	454 GB
13 / 82	Qwen3.5 397B A17B243 GB · 9 benchmarks		–	–	112	–	115	9	243 GB
14 / 84	Kimi K2.5618 GB · 20 benchmarks		110	111	–	117	114	20	618 GB
15 / 85	MiniMax M2.5140 GB · 11 benchmarks		109	110	–	–	114	11	140 GB
16 / 86	DeepSeek V3.2 Exp413 GB · 10 benchmarks		107	–	–	101	113	10	413 GB
17 / 88	Qwen3.5 122B A10B75.5 GB · 6 benchmarks		–	–	–	96	112	6	75.5 GB
18 / 90	DeepSeek V3.2413 GB · 13 benchmarks		–	108	–	–	112	13	413 GB
19 / 93	DeepSeek V4 Flash175 GB · 10 benchmarks		107	–	–	111	112	10	175 GB
20 / 96	MiMo V2.5 Pro616 GB · 6 benchmarks		116	–	–	115	111	6	616 GB
21 / 99	Qwen3.5 27B17.9 GB · 6 benchmarks		98	–	–	–	111	6	17.9 GB
22 / 103	Gemma 4 31B IT21.7 GB · 11 benchmarks		108	–	–	105	109	11	21.7 GB
23 / 104	GLM 5454 GB · 15 benchmarks		110	118	–	–	109	15	454 GB
24 / 105	DeepSeek V3.1 Terminus412 GB · 5 benchmarks		107	–	–	99	108	5	412 GB
25 / 107	MiniMax M3258 GB · 19 benchmarks		113	113	106	118	107	19	258 GB
26 / 109	Qwen3.6 35B A3B22.0 GB · 12 benchmarks		105	–	109	110	107	12	22.0 GB
27 / 110	Qwen3.6 27B17.9 GB · 12 benchmarks		102	–	114	111	106	12	17.9 GB
28 / 114	Qwen3 235B A22B Thinking 2507143 GB · 13 benchmarks		102	–	–	108	106	13	143 GB
29 / 115	Qwen3.5 35B A3B22.0 GB · 8 benchmarks		–	–	–	108	106	8	22.0 GB
30 / 116	DeepSeek V3.1412 GB · 7 benchmarks		–	–	–	–	105	7	412 GB
31 / 119	GLM 4.7218 GB · 13 benchmarks		101	106	111	111	104	13	218 GB
32 / 125	Gemma 4 26B A4B IT16.5 GB · 9 benchmarks		104	–	–	102	101	9	16.5 GB
33 / 127	DeepSeek R1 0528412 GB · 10 benchmarks		107	–	–	–	100	10	412 GB
34 / 129	Qwen3 235B A22B143 GB · 9 benchmarks		101	–	–	–	99	9	143 GB
35 / 131	Qwen3 235B A22B Instruct 2507143 GB · 8 benchmarks		102	–	–	–	98	8	143 GB
36 / 132	Qwen3.5 9B6.6 GB · 9 benchmarks		–	–	–	105	97	9	6.6 GB
37 / 134	GPT OSS 120B70.8 GB · 20 benchmarks		98	82	–	103	97	20	70.8 GB
38 / 136	DeepSeek R1412 GB · 13 benchmarks		100	–	–	99	96	13	412 GB
39 / 138	Mistral Small 4 119B 260381.6 GB · 3 benchmarks		–	–	–	–	96	3	81.6 GB
40 / 142	Qwen3 30B A3B Thinking 250719.4 GB · 8 benchmarks		–	–	–	99	94	8	19.4 GB
41 / 144	Qwen3 30B A3B Instruct 250719.4 GB · 5 benchmarks		–	–	–	–	92	5	19.4 GB
42 / 150	Qwen3 32B22.6 GB · 9 benchmarks		93	–	–	96	91	9	22.6 GB
43 / 151	GPT OSS 20B13.1 GB · 11 benchmarks		101	–	–	92	91	11	13.1 GB
44 / 153	Kimi K2 Instruct617 GB · 10 benchmarks		102	102	–	–	90	10	617 GB
45 / 155	Mistral Large 3 675B Instruct 2512454 GB · 5 benchmarks		93	–	–	97	89	5	454 GB
46 / 156	Magistral Small 250617.8 GB · 4 benchmarks		–	–	–	–	88	4	17.8 GB
47 / 158	Qwen3 14B11.9 GB · 8 benchmarks		–	–	–	94	88	8	11.9 GB
48 / 159	DeepSeek v3 0324412 GB · 12 benchmarks		100	–	–	98	88	12	412 GB
49 / 167	Qwen2.5 72B Instruct49.3 GB · 8 benchmarks		–	–	–	–	86	8	49.3 GB
50 / 168	Llama 4 Maverick 17B 128E Instruct273 GB · 16 benchmarks		81	–	–	96	85	16	273 GB
51 / 169	Llama 3.1 405B Instruct276 GB · 7 benchmarks		–	–	–	–	85	7	276 GB
52 / 171	Magistral Small 250918.1 GB · 6 benchmarks		–	–	–	83	85	6	18.1 GB
53 / 173	Mistral Large Instruct 241179.8 GB · 5 benchmarks		–	–	–	–	84	5	79.8 GB
54 / 175	C4ai Command A 03 202578.7 GB · 3 benchmarks		–	–	–	–	84	3	78.7 GB
55 / 176	Qwen3 30B A3B19.4 GB · 7 benchmarks		–	–	–	–	83	7	19.4 GB
56 / 177	Llama 3.1 70B Instruct51.3 GB · 7 benchmarks		–	–	–	–	83	7	51.3 GB
57 / 178	Mistral Large Instruct 240786.4 GB · 6 benchmarks		–	–	–	–	83	6	86.4 GB
58 / 179	Mistral Small 3.2 24B Instruct 250618.1 GB · 6 benchmarks		–	–	–	82	83	6	18.1 GB
59 / 181	Llama 3.3 70B Instruct51.3 GB · 12 benchmarks		76	–	–	81	82	12	51.3 GB
60 / 182	Mistral Small 3.1 24B Instruct 250318.1 GB · 6 benchmarks		–	–	–	81	82	6	18.1 GB
61 / 183	DeepSeek v3412 GB · 8 benchmarks		97	–	–	90	81	8	412 GB
62 / 184	Qwen3 8B7.6 GB · 8 benchmarks		–	–	–	88	81	8	7.6 GB
63 / 185	Llama 4 Scout 17B 16E Instruct77.0 GB · 11 benchmarks		–	–	–	83	80	11	77.0 GB
64 / 188	Mixtral 8x22B Instruct v0.188.4 GB · 5 benchmarks		–	–	–	–	76	5	88.4 GB
65 / 189	C4ai Command R Plus 08 202473.8 GB · 3 benchmarks	—	–	–	–	–	74	3	73.8 GB
66 / 190	Gemma 3 27B IT21.7 GB · 10 benchmarks		70	–	–	81	74	10	21.7 GB
67 / 194	Llama 3.1 8B Instruct7.9 GB · 10 benchmarks		56	–	–	56	69	10	7.9 GB
68 / 195	Mixtral 8x7B Instruct v0.130.5 GB · 4 benchmarks		–	–	–	–	69	4	30.5 GB
69 / 196	Gemma 3 4B IT5.0 GB · 6 benchmarks		–	–	–	–	69	6	5.0 GB
70 / 199	Gemma 3 12B IT10.9 GB · 8 benchmarks		–	–	–	71	66	8	10.9 GB
71 / 201	Qwen2.5 7B Instruct5.8 GB · 6 benchmarks		–	–	–	–	65	6	5.8 GB
72 / 203	C4ai Command R 08 202425.1 GB · 2 benchmarks	—	–	–	–	–	64	2	25.1 GB
73 / 204	Meta Llama 3 8B Instruct7.9 GB · 6 benchmarks		–	–	–	–	64	6	7.9 GB
74 / 205	Llama 2 13B Chat HF11.5 GB · 2 benchmarks	—	–	–	–	–	57	2	11.5 GB
75 / 207	Llama 2 70B Chat HF50.2 GB · 4 benchmarks		–	–	–	–	45	4	50.2 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.