llmrun
Math

LiveBench Math Leaderboard

LiveBench Math measures mathematical problem-solving on contamination-free, regularly-refreshed questions, including competition-style problems.

Source: livebench19 open models ranked+51 proprietaryData through Jun 2026

Open modelsAll models

All models ranked on LiveBench Math

Proprietary / closed models are shown dimmed — you can't run them locally, but they show where the open field stands.

#ModelScoreVRAM @ Q4
1Claude Opus 5.5 Max · proprietary
97.1
API only
2Claude Fable 5.1 Max · proprietary
97.0
API only
3GPT 6.1 Sol Max · proprietary
96.8
API only
4GPT 6 Astra Max · proprietary
96.8
API only
5Claude Opus 5.5 · proprietary
96.8
API only
6Claude Sonnet 5.5 · proprietary
96.7
API only
7GPT 6.1 Sol · proprietary
96.5
API only
8Muse Wonder 1.0 Max · proprietary
96.5
API only
9GPT 6 Sol Max · proprietary
96.4
API only
10GPT 5.6 Sol Max · proprietary
96.2
API only
11Claude Sonnet 5.5 Max · proprietary
96.1
API only
12Claude Fable 5 Max · proprietary
96.0
API only
13Muse Spark 1.3 · proprietary
96.0
API only
14GPT 5.5 · proprietary
95.9
API only
15Claude Opus 5 Max · proprietary
95.7
API only
16Grok 4.7 · proprietary
95.7
API only
17DeepSeek V4 Pro 0813 · 1650.5B
95.1
~991 GB
18GPT 5.6 Terra Max · proprietary
94.9
API only
19Claude Opus 4.8 Max · proprietary
94.3
API only
20GPT 5.4 · proprietary
94.2
API only
21Mistral Large 4 · proprietary
93.6
API only
22Gemini 3.7 Flash · proprietary
93.5
API only
23DeepSeek V4.1 Flash · 763.2B
93.3
~458 GB
24GPT 5.2 · proprietary
93.2
API only
25Claude Sonnet 5 · proprietary
92.9
API only
26Claude Opus 4.7 · proprietary
92.8
API only
27Claude Haiku 5.5 · proprietary
92.8
API only
28Grok 4.6 · proprietary
92.6
API only
29Gemini 3.8 Flash · proprietary
91.6
API only
30Qwen3.8 Max · proprietary
91.3
API only
31Muse Spark 1.2 · proprietary
91.2
API only
32Gemini 3.1 Pro · proprietary
91.0
API only
33GPT 5.4 Nano · proprietary
91.0
API only
34Grok 4.5 · proprietary
90.8
API only
35DeepSeek V4 Pro · 1598.8B
90.7
~960 GB
36Claude Opus 4.5 · proprietary
90.4
API only
37Smaug Flash · 304.2B
90.1
~609 GB
38GLM 5.2 · 753.3B
89.8
~456 GB
39Smaug Mini · proprietary
89.7
API only
40Claude Opus 4.6 · proprietary
89.3
API only
41GPT 6 Luna Max · proprietary
89.1
API only
42GPT 5.2 Codex · proprietary
88.8
API only
43NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 560.5B
88.7
~370 GB
44Inkling · 952.4B
88.4
~572 GB
45Gemini 3.5 Flash · proprietary
88.2
API only
46GLM 5.3 · 753.3B
87.9
~456 GB
47DeepSeek V4 Flash Vision Exp · 304.6B
87.8
~183 GB
48GPT 5.6 Luna Max · proprietary
87.2
API only
49Muse Spark 1.1 · proprietary
87.1
API only
50Claude Sonnet 4.6 · proprietary
87.0
API only
51DeepSeek V4 Flash 0731 · 304.2B
86.8
~183 GB
52Claude Haiku 5.5 Max · proprietary
86.4
API only
53Gemini 3.6 Flash · proprietary
86.4
API only
54Qwen3.8 27B · 27.8B
86.2
~17 GB
55Qwen3.8 Flash Next · 180.0B
85.8
~108 GB
56Qwen3.7 Max · proprietary
85.3
API only
57Kimi K3 · 2779.9B
84.4
~1674 GB
58Grok 4.3 · proprietary
84.3
API only
59Kimi K2.6 · 1026.9B
84.3
~620 GB
60Smaug Agentic · proprietary
83.9
API only
61Qwen3.6 Plus · proprietary
83.7
API only
62GLM 5.3 Flash · 321.3B
81.2
~195 GB
63Qwen3.6 27B · 27.8B
79.9
~17 GB
64DeepSeek V4 Flash · 290.9B
79.7
~175 GB
65Kimi K2.7 Code · 1026.9B
79.6
~620 GB
66GPT 5.4 Mini · proprietary
78.5
API only
67Grok Build 0.1 · proprietary
78.4
API only
68Ox Alpha Max · proprietary
77.5
API only
69MiniMax M3 · 427.0B
77.0
~257 GB
70Gemini 3.5 Flash Lite · proprietary
73.7
API only

Score vs model size

Which models give the most quality for their size — the ones worth running locally.

100B1Tmodel size (log scale) →95.177.0DeepSeek V4 Pro · 1.6T · 90.7GLM 5.2 · 753B · 89.8NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 561B · 88.7Inkling · 952B · 88.4GLM 5.3 · 753B · 87.9DeepSeek V4 Flash Vision Exp · 305B · 87.8DeepSeek V4 Flash 0731 · 304B · 86.8Qwen3.8 Flash Next · 180B · 85.8Kimi K3 · 2.8T · 84.4Kimi K2.6 · 1T · 84.3GLM 5.3 Flash · 321B · 81.2Qwen3.6 27B · 28B · 79.9DeepSeek V4 Flash · 291B · 79.7Kimi K2.7 Code · 1T · 79.6MiniMax M3 · 427B · 77.0Qwen3.8 27B · 28B · 86.2Qwen3.8 27BSmaug Flash · 304B · 90.1Smaug FlashDeepSeek V4.1 Flash · 763B · 93.3DeepSeek V4.1 FlashDeepSeek V4 Pro 0813 · 1.7T · 95.1DeepSeek V4 Pro 0813
Each dot is a model. Up = higher score, left = smaller (easier to run locally). The dashed line marks the efficiency frontier — the best score you can get at each size or smaller.
  • Qwen3.8 27B, 28B, score 86.2 — on the efficiency frontier (best score at its size or smaller).
  • Smaug Flash, 304B, score 90.1 — on the efficiency frontier (best score at its size or smaller).
  • DeepSeek V4.1 Flash, 763B, score 93.3 — on the efficiency frontier (best score at its size or smaller).
  • DeepSeek V4 Pro 0813, 1.7T, score 95.1 — on the efficiency frontier (best score at its size or smaller).

LiveBench Math: frequently asked questions

What is the best open LLM on LiveBench Math?
DeepSeek V4 Pro 0813 is the top open model on LiveBench Math, scoring 95.1. Among all models tested — including proprietary ones — it ranks #17. The top model overall is Claude Opus 5.5 Max (Anthropic) at 97.1.
What's the best LiveBench Math model you can run on a 24 GB GPU?
Qwen3.8 27B is the highest-scoring open model that fits in 24 GB at 4-bit quantization (~17 GB), scoring 86.2 on LiveBench Math.
Can open models match proprietary models on LiveBench Math?
Not quite on LiveBench Math: the strongest proprietary model (Claude Opus 5.5 Max) scores 97.1, ahead of the best open model (DeepSeek V4 Pro 0813) at 95.1 — but you can run the open one yourself.

Scores aggregated from livebench. llmrun does not run this benchmark — see the source for methodology, or the about benchmarks for what it measures.