Best open LLMs for math

Open models ranked by mathematics, from competition and research math boards such as AIME and FrontierMath. The score shown is the overall llmrun Score; the highlighted column is the mathematics 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 math llmrun Score
#	Model	llmrun Score	Coding	Agents	Math	Science	Reasoning	Benchmarks	VRAM
1	Kimi K31674 GB · 26 benchmarks		73	61	81	52	63	26	1674 GB
2	DeepSeek V4.1 Flash458 GB · 10 benchmarks		54	–	76	45	57	10	458 GB
3	GLM 5.3456 GB · 21 benchmarks		67	68	72	49	58	21	456 GB
4	DeepSeek V4 Pro 0813991 GB · 21 benchmarks		63	–	71	46	65	21	991 GB
5	DeepSeek V4 Flash 0731183 GB · 21 benchmarks		59	–	71	45	62	21	183 GB
6	GLM 5.2456 GB · 24 benchmarks		58	60	68	47	52	24	456 GB
7	Kimi K2.6620 GB · 18 benchmarks		57	–	63	43	51	18	620 GB
8	GLM 5.3 Flash195 GB · 17 benchmarks		52	63	62	44	37	17	195 GB
9	Kimi K2.7 Code620 GB · 19 benchmarks		53	46	61	40	50	19	620 GB
10	DeepSeek V4 Pro960 GB · 22 benchmarks		49	–	57	44	49	22	960 GB
11	Qwen3.8 27B17.4 GB · 9 benchmarks		51	–	57	37	50	9	17.4 GB
12	GLM 5.1457 GB · 13 benchmarks		53	54	57	40	46	13	457 GB
13	Inkling Small160 GB · 12 benchmarks		–	–	57	41	54	12	160 GB
14	Qwen3.6 27B17.4 GB · 12 benchmarks		42	–	53	36	37	12	17.4 GB
15	Qwen3.5 397B A17B266 GB · 9 benchmarks		–	–	52	–	46	9	266 GB
16	Inkling572 GB · 21 benchmarks		43	–	52	40	53	21	572 GB
17	NVIDIA Nemotron 3 Ultra 550B A55B BF16370 GB · 14 benchmarks		43	–	52	37	46	14	370 GB
18	GLM 4.7216 GB · 13 benchmarks		41	32	51	36	35	13	216 GB
19	Qwen3.6 35B A3B22.0 GB · 12 benchmarks		44	–	49	35	38	12	22.0 GB
20	MiniMax M3257 GB · 19 benchmarks		52	40	46	41	38	19	257 GB
21	DeepSeek R1 0528415 GB · 11 benchmarks		46	–	44	–	32	11	415 GB
22	DeepSeek R1415 GB · 14 benchmarks		39	–	40	28	29	14	415 GB
23	DeepSeek R1 Distill Llama 70B43.3 GB · 3 benchmarks	—	–	–	39	–	–	3	43.3 GB
24	DeepSeek R1 Distill Qwen 14B9.6 GB · 4 benchmarks		–	–	38	–	–	4	9.6 GB
25	DeepSeek v3 0324415 GB · 13 benchmarks		39	–	33	27	24	13	415 GB
26	Gemma 3 27B IT18.1 GB · 11 benchmarks		18	–	29	16	17	11	18.1 GB
27	Llama 4 Maverick 17B 128E Instruct265 GB · 17 benchmarks		25	–	29	26	23	17	265 GB
28	DeepSeek v3415 GB · 10 benchmarks		37	–	26	22	21	10	415 GB
29	Phi 49.5 GB · 6 benchmarks		–	–	26	–	–	6	9.5 GB
30	Qwen2.5 72B Instruct44.6 GB · 9 benchmarks		–	–	25	–	23	9	44.6 GB
31	Llama 4 Scout 17B 16E Instruct71.7 GB · 12 benchmarks		–	–	24	18	20	12	71.7 GB
32	Qwen2.5 32B Instruct20.5 GB · 5 benchmarks		–	–	23	–	–	5	20.5 GB
33	Llama 3.1 405B Instruct268 GB · 8 benchmarks		–	–	22	–	22	8	268 GB
34	Mistral Large Instruct 241174.6 GB · 6 benchmarks		–	–	22	–	22	6	74.6 GB
35	Mistral Small 3.1 24B Instruct 250315.1 GB · 7 benchmarks		–	–	21	17	21	7	15.1 GB
36	Mistral Large Instruct 240780.9 GB · 7 benchmarks		–	–	20	–	22	7	80.9 GB
37	Mistral Small 24B Instruct 250114.9 GB · 4 benchmarks		–	–	20	–	–	4	14.9 GB
38	Llama 3.1 Tulu 3 70B DPO43.3 GB · 3 benchmarks	—	–	–	19	–	–	3	43.3 GB
39	Llama 3.3 70B Instruct46.6 GB · 13 benchmarks		22	–	19	17	21	13	46.6 GB
40	Llama 3.2 90B Vision Instruct58.5 GB · 4 benchmarks		–	–	18	–	–	4	58.5 GB
41	Llama 3.1 70B Instruct46.6 GB · 8 benchmarks		–	–	18	–	22	8	46.6 GB
42	Gemma 2 27B IT18.0 GB · 4 benchmarks		–	–	16	–	–	4	18.0 GB
43	Meta Llama 3 70B Instruct46.6 GB · 5 benchmarks		–	–	14	–	–	5	46.6 GB
44	Hermes 2 Theta Llama 3 70B43.3 GB · 3 benchmarks	—	–	–	14	–	–	3	43.3 GB
45	Llama 3.1 8B Instruct5.3 GB · 11 benchmarks		14	–	14	7	16	11	5.3 GB
46	Gemma 2 9B IT6.1 GB · 4 benchmarks		–	–	14	–	–	4	6.1 GB
47	Deepseek Llm 67B Chat41.3 GB · 4 benchmarks		–	–	10	–	–	4	41.3 GB
48	Meta Llama 3 8B Instruct5.3 GB · 7 benchmarks		–	–	10	–	14	7	5.3 GB
49	Mistral 7B Instruct v0.34.9 GB · 4 benchmarks		–	–	10	–	–	4	4.9 GB
50	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.