Best open LLMs for reasoning

Open models ranked by reasoning, from hard multi-step reasoning boards such as ARC-AGI and HLE. The score shown is the overall llmrun Score; the highlighted column is the reasoning skill score used for the order.

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 reasoning llmrun Score
#	Model	llmrun Score	Coding	Agents	Math	Science	Reasoning	Benchmarks	VRAM
1	DeepSeek V4 Pro 0813991 GB · 21 benchmarks		63	–	71	46	65	21	991 GB
2	Kimi K31674 GB · 26 benchmarks		73	61	81	52	63	26	1674 GB
3	DeepSeek V4 Flash 0731183 GB · 21 benchmarks		59	–	71	45	62	21	183 GB
4	GLM 5.3456 GB · 21 benchmarks		67	68	72	49	58	21	456 GB
5	DeepSeek V4.1 Flash458 GB · 10 benchmarks		54	–	76	45	57	10	458 GB
6	Inkling Small160 GB · 12 benchmarks		–	–	57	41	54	12	160 GB
7	Inkling572 GB · 21 benchmarks		43	–	52	40	53	21	572 GB
8	GLM 5.2456 GB · 24 benchmarks		58	60	68	47	52	24	456 GB
9	Kimi K2.6620 GB · 18 benchmarks		57	–	63	43	51	18	620 GB
10	Qwen3.8 27B17.4 GB · 9 benchmarks		51	–	57	37	50	9	17.4 GB
11	Kimi K2.7 Code620 GB · 19 benchmarks		53	46	61	40	50	19	620 GB
12	DeepSeek V4 Pro960 GB · 22 benchmarks		49	–	57	44	49	22	960 GB
13	NVIDIA Nemotron 3 Ultra 550B A55B BF16370 GB · 14 benchmarks		43	–	52	37	46	14	370 GB
14	GLM 5.1457 GB · 13 benchmarks		53	54	57	40	46	13	457 GB
15	Qwen3.5 397B A17B266 GB · 9 benchmarks		–	–	52	–	46	9	266 GB
16	Kimi K2.5620 GB · 20 benchmarks		49	37	–	40	45	20	620 GB
17	MiniMax M2.5138 GB · 11 benchmarks		47	37	–	–	45	11	138 GB
18	DeepSeek V3.2 Exp415 GB · 10 benchmarks		46	–	–	30	44	10	415 GB
19	Qwen3.5 122B A10B82.6 GB · 6 benchmarks		–	–	–	26	43	6	82.6 GB
20	DeepSeek V3.2415 GB · 13 benchmarks		35	36	–	–	43	13	415 GB
21	DeepSeek V4 Flash175 GB · 10 benchmarks		46	–	–	36	43	10	175 GB
22	MiMo V2.5 Pro615 GB · 6 benchmarks		54	–	–	38	42	6	615 GB
23	Qwen3.5 27B17.4 GB · 6 benchmarks		38	–	–	–	41	6	17.4 GB
24	Gemma 4 31B IT20.4 GB · 11 benchmarks		47	–	–	32	40	11	20.4 GB
25	GLM 5457 GB · 15 benchmarks		49	45	–	–	40	15	457 GB
26	DeepSeek V3.1 Terminus415 GB · 5 benchmarks		47	–	–	31	39	5	415 GB
27	MiniMax M3257 GB · 19 benchmarks		52	40	46	41	38	19	257 GB
28	Qwen3.6 35B A3B22.0 GB · 12 benchmarks		44	–	49	35	38	12	22.0 GB
29	Qwen3.6 27B17.4 GB · 12 benchmarks		42	–	53	36	37	12	17.4 GB
30	GLM 5.3 Flash195 GB · 17 benchmarks		52	63	62	44	37	17	195 GB
31	Qwen3 235B A22B Thinking 2507142 GB · 13 benchmarks		42	–	–	34	37	13	142 GB
32	Qwen3.5 35B A3B22.0 GB · 8 benchmarks		–	–	–	34	37	8	22.0 GB
33	DeepSeek V3.1415 GB · 7 benchmarks		–	–	–	–	36	7	415 GB
34	GLM 4.7216 GB · 13 benchmarks		41	32	51	36	35	13	216 GB
35	Gemma 4 26B A4B IT16.1 GB · 9 benchmarks		43	–	–	30	33	9	16.1 GB
36	DeepSeek R1 0528415 GB · 11 benchmarks		46	–	44	–	32	11	415 GB
37	Qwen3 235B A22B142 GB · 10 benchmarks		41	–	–	–	31	10	142 GB
38	Qwen3 235B A22B Instruct 2507142 GB · 8 benchmarks		41	–	–	–	31	8	142 GB
39	Qwen3.5 9B6.4 GB · 9 benchmarks		–	–	–	32	30	9	6.4 GB
40	GPT OSS 120B70.5 GB · 20 benchmarks		37	14	–	31	30	20	70.5 GB
41	DeepSeek R1415 GB · 14 benchmarks		39	–	40	28	29	14	415 GB
42	Mistral Small 4 119B 260373.2 GB · 3 benchmarks	—	–	–	–	–	29	3	73.2 GB
43	Qwen3 30B A3B Thinking 250718.7 GB · 8 benchmarks		–	–	–	28	27	8	18.7 GB
44	Qwen3 30B A3B Instruct 250718.7 GB · 5 benchmarks		–	–	–	–	26	5	18.7 GB
45	Qwen3 32B20.3 GB · 9 benchmarks		33	–	–	26	26	9	20.3 GB
46	GPT OSS 20B12.9 GB · 11 benchmarks		40	–	–	24	26	11	12.9 GB
47	Kimi K2 Instruct620 GB · 10 benchmarks		41	28	–	–	25	10	620 GB
48	Mistral Large 3 675B Instruct 2512446 GB · 5 benchmarks		33	–	–	26	25	5	446 GB
49	Magistral Small 250614.9 GB · 4 benchmarks		–	–	–	–	24	4	14.9 GB
50	DeepSeek v3 0324415 GB · 13 benchmarks		39	–	33	27	24	13	415 GB
51	Qwen3 14B9.5 GB · 8 benchmarks		–	–	–	25	24	8	9.5 GB
52	Qwen2.5 72B Instruct44.6 GB · 9 benchmarks		–	–	25	–	23	9	44.6 GB
53	Llama 4 Maverick 17B 128E Instruct265 GB · 17 benchmarks		25	–	29	26	23	17	265 GB
54	Llama 3.1 405B Instruct268 GB · 8 benchmarks		–	–	22	–	22	8	268 GB
55	Magistral Small 250915.1 GB · 6 benchmarks		–	–	–	18	22	6	15.1 GB
56	Mistral Large Instruct 241174.6 GB · 6 benchmarks		–	–	22	–	22	6	74.6 GB
57	C4ai Command A 03 202573.3 GB · 3 benchmarks	—	–	–	–	–	22	3	73.3 GB
58	Qwen3 30B A3B18.7 GB · 7 benchmarks		–	–	–	–	22	7	18.7 GB
59	Llama 3.1 70B Instruct46.6 GB · 8 benchmarks		–	–	18	–	22	8	46.6 GB
60	Mistral Large Instruct 240780.9 GB · 7 benchmarks		–	–	20	–	22	7	80.9 GB
61	Mistral Small 3.2 24B Instruct 250615.1 GB · 6 benchmarks		–	–	–	18	22	6	15.1 GB
62	Llama 3.3 70B Instruct46.6 GB · 13 benchmarks		22	–	19	17	21	13	46.6 GB
63	Mistral Small 3.1 24B Instruct 250315.1 GB · 7 benchmarks		–	–	21	17	21	7	15.1 GB
64	DeepSeek v3415 GB · 10 benchmarks		37	–	26	22	21	10	415 GB
65	Qwen3 8B5.5 GB · 8 benchmarks		–	–	–	21	21	8	5.5 GB
66	Llama 4 Scout 17B 16E Instruct71.7 GB · 12 benchmarks		–	–	24	18	20	12	71.7 GB
67	Mixtral 8x22B Instruct v0.185.2 GB · 6 benchmarks		–	–	–	–	19	6	85.2 GB
68	C4ai Command R Plus 08 202468.5 GB · 3 benchmarks	—	–	–	–	–	18	3	68.5 GB
69	Gemma 3 27B IT18.1 GB · 11 benchmarks		18	–	29	16	17	11	18.1 GB
70	Llama 3.1 8B Instruct5.3 GB · 11 benchmarks		14	–	14	7	16	11	5.3 GB
71	Mixtral 8x7B Instruct v0.128.6 GB · 5 benchmarks		–	–	–	–	16	5	28.6 GB
72	Gemma 3 4B IT2.8 GB · 6 benchmarks		–	–	–	–	16	6	2.8 GB
73	Gemma 3 12B IT8.0 GB · 8 benchmarks		–	–	–	12	15	8	8.0 GB
74	Qwen2.5 7B Instruct5.0 GB · 6 benchmarks		–	–	–	–	15	6	5.0 GB
75	C4ai Command R 08 202421.3 GB · 2 benchmarks	—	–	–	–	–	14	2	21.3 GB
76	Meta Llama 3 8B Instruct5.3 GB · 7 benchmarks		–	–	10	–	14	7	5.3 GB
77	Llama 2 13B Chat HF8.6 GB · 2 benchmarks	—	–	–	–	–	12	2	8.6 GB
78	Llama 2 70B Chat HF45.5 GB · 5 benchmarks		–	–	9	–	10	5	45.5 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 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.