Leaderboards

Benchmarks

Every score carries its origin — leaderboard is independent,vendor is self-reported. History is kept, never overwritten.

Reasoning280Coding266Math174Agentic110Chat515Speech654
Arena Hard Prompts143GPQA Diamond16LiveBench35LiveBench Data Analysis35LiveBench Reasoning35MMLU0MMLU-Pro16

MMLU-Pro

% correct, 12k harder multi-task questionshigher is better
Provenanceindependent leaderboard16 modelsscores Aug 3, 2026
1Qwen2.5 72B InstructQwen72.7B51.3128~1.1Aug 3, 2026leaderboard2Mistral LargeMistral AI50.78Aug 3, 2026leaderboard3Llama 3.3 70B InstructMeta70.6B48.1150~1.1Aug 3, 2026leaderboard4Llama 3.1 70B InstructMeta70.6B47.9120~1.1Aug 3, 2026leaderboard5Phi 4Microsoft14.7B47.63405Aug 3, 2026leaderboard6R1 Distill Llama 70BDeepSeek70.6B41.652~0.93Aug 3, 2026leaderboard7Hermes 3 70B InstructNous Research70.6B41.459~0.93Aug 3, 2026leaderboard8Mixtral 8x22B InstructMistral AI141B38.76~0.44Aug 3, 2026leaderboard9Gemma 2 27BGoogle27.2B38.359~2.2Aug 3, 2026leaderboard10Qwen2.5 Coder 32B InstructQwen32.8B37.938~1.8Aug 3, 2026leaderboard11Qwen2.5 7B InstructQwen7.6B37.8540~7.9Aug 3, 2026leaderboard12Command R7B (12-2024)Cohere28.6191Aug 3, 2026leaderboard13Mistral NemoMistral AI12.2B28933~3.6Aug 3, 2026leaderboard14Llama 3.1 8B InstructMeta8B25.1628~5Aug 3, 2026leaderboard15Llama 3.2 3B InstructMeta3.2B20.2404~10Aug 3, 2026leaderboard16Llama 3.2 1B InstructMeta1.2B9.3465~12Aug 3, 2026leaderboard

A tilde in Score / GB VRAM means the memory figure behind it is calculated from the parameter count rather than read from a published file. Those ratios are shown to two significant figures, and rows within a few percent of each other are not distinguishable.