Arena Hard

AI benchmark results and model performance

Type: percentage

Recorded Results

Average Score

63.25%

Score Range

26.70% - 95.60%

Leading Model (lowest score)

26.70% - Granite 4.0 Micro

Scores Over Time

Individual benchmark scores plotted by date.

Models Using This Benchmark

Organisation	Model	Reported	Top Score	Info	Self Reported	Source
IBM	Granite 4.0 Micro	02 Oct 2025	26.70%	inferred high-confidence family alias from granite-4.0-tiny-preview (score=0.4700; benches=12)	Yes	Source
IBM	Granite 4.0 Small	02 Oct 2025	26.70%	inferred high-confidence family alias from granite-4.0-tiny-preview (score=0.4700; benches=12)	Yes	Source
IBM	Granite 4.0 Tiny	02 Oct 2025	26.70%	inferred alias from granite-4.0-tiny-preview	Yes	Source
IBM	Granite 4.0 Tiny Preview	02 May 2025	26.70%	-	Yes	Source
Mistral	Ministral 3.0 3B	02 Dec 2025	30.50%	-	Yes	Source
Microsoft	Phi 4 Mini	01 Feb 2025	32.80%	-	Yes	Source
Microsoft	Phi 3 Mini 128K Instruct	-	37%	inferred family alias from phi-3.5-mini-instruct (score=0.3533; benches=31)	Yes	Source
Microsoft	Phi 3.5 mini instruct	23 Aug 2024	37%	-	Yes	Source
IBM	Granite 4.1 3B	29 Apr 2026	37.80%	-	Yes	Source
Microsoft	Phi 3.5 MoE instruct	23 Aug 2024	37.90%	-	Yes	Source
Mistral	Ministral 3.0 8B	02 Dec 2025	50.90%	-	Yes	Source
Qwen	Qwen 7B	-	52%	inferred family alias from qwen-2.5-7b-instruct (score=0.3083; benches=14)	Yes	Source
Mistral	Ministral 3.0 14B	02 Dec 2025	55.10%	-	Yes	Source
IBM	Granite 3.0 8B Instruct	-	57.56%	inferred high-confidence family alias from granite-3.3-8b-instruct (score=0.4911; benches=14)	Yes	Source
IBM	Granite 3.1 8B Instruct	-	57.56%	inferred high-confidence family alias from granite-3.3-8b-instruct (score=0.4911; benches=14)	Yes	Source
IBM	Granite 3.2 8B Instruct	-	57.56%	inferred high-confidence family alias from granite-3.3-8b-instruct (score=0.4911; benches=14)	Yes	Source
IBM	Granite 3.2 8B Instruct Preview	-	57.56%	inferred high-confidence family alias from granite-3.3-8b-instruct (score=0.4687; benches=14)	Yes	Source
IBM	Granite 3.3 2B Instruct	16 Apr 2025	57.56%	inferred family alias from granite-3.3-8b-instruct (score=0.3627; benches=14)	Yes	Source
IBM	Granite 3.3 8B Instruct	16 Apr 2025	57.56%	-	Yes	Source
IBM	Granite Guardian 3.0 8B	-	57.56%	inferred family alias from granite-3.3-8b-instruct (score=0.4062; benches=14)	Yes	Source
IBM	Granite Guardian 3.1 8B	-	57.56%	inferred family alias from granite-3.3-8b-instruct (score=0.4062; benches=14)	Yes	Source
IBM	Granite Guardian 3.3 8B	-	57.56%	inferred high-confidence family alias from granite-3.3-8b-instruct (score=0.5071; benches=14)	Yes	Source
IBM	Granite Speech 3.2 8B	-	57.56%	inferred family alias from granite-3.3-8b-instruct (score=0.4062; benches=14)	Yes	Source
IBM	Granite Speech 3.3 8B	-	57.56%	inferred high-confidence family alias from granite-3.3-8b-instruct (score=0.5071; benches=14)	Yes	Source
Mistral	Mistral Small 1.0	26 Feb 2024	58.30%	inferred family alias from mistral-small-latest (score=0.3650; benches=9)	Yes	Source
Mistral	Mistral Small 2.0	17 Sept 2024	58.30%	inferred family alias from mistral-small-latest (score=0.3650; benches=9)	Yes	Source
Mistral	Mistral Small 4	16 Mar 2026	58.30%	-	Yes	Source
Mistral	Mistral Small Creative	16 Dec 2025	58.30%	inferred family alias from mistral-small-latest (score=0.4273; benches=9)	Yes	Source
IBM	Granite 4.1 8B	29 Apr 2026	68.98%	-	Yes	Source
IBM	Granite 4.1 30B	29 Apr 2026	71.02%	-	Yes	Source
Microsoft	Phi 4 Reasoning	30 Apr 2025	73.30%	-	Yes	Source
Microsoft	Phi 1	-	75.40%	inferred family alias from phi-4 (score=0.3100; benches=13)	Yes	Source
Microsoft	Phi 2	-	75.40%	inferred family alias from phi-4 (score=0.3100; benches=13)	Yes	Source
Microsoft	Phi 4	12 Dec 2024	75.40%	-	Yes	Source
DeepSeek	DeepSeek V2.5 (2024-09-05)	05 Sept 2024	76.20%	inferred alias from deepseek-v2.5	Yes	Source
DeepSeek	DeepSeek V2.5 (2024-12-10)	10 Dec 2024	76.20%	inferred alias from deepseek-v2.5	Yes	Source
Microsoft	Phi 4 Reasoning Plus	30 Apr 2025	79%	-	Yes	Source
Qwen	Qwen 72B	-	81.20%	inferred family alias from qwen-2.5-72b-instruct (score=0.3060; benches=14)	Yes	Source
Nvidia	Llama 3.3 Nemotron Super 49B v1	18 Mar 2025	88.30%	-	Yes	Source
Nvidia	Llama 3.3 Nemotron Super 49B V1.5	-	88.30%	inferred version-family alias from llama-3.3-nemotron-super-49b-v1	Yes	Source
Qwen	Qwen 3 30B A3B	-	91%	-	Yes	Source
Qwen	Qwen 3 30B A3B Instruct 2507	-	91%	inferred version-family alias from qwen3-30b-a3b	Yes	Source
Qwen	Qwen 3 30B A3B Thinking 2507	-	91%	inferred version-family alias from qwen3-30b-a3b	Yes	Source
Qwen	Qwen 3 Coder 30B A3B Instruct	-	91%	inferred high-confidence family alias from qwen3-30b-a3b (score=0.5007; benches=8)	Yes	Source
Qwen	Qwen 3 Omni 30B A3B Captioner	-	91%	inferred family alias from qwen3-30b-a3b (score=0.4129; benches=8)	Yes	Source
Qwen	Qwen 3 Omni 30B A3B Instruct	-	91%	inferred high-confidence family alias from qwen3-30b-a3b (score=0.4819; benches=8)	Yes	Source
Qwen	Qwen 3 Omni 30B A3B Thinking	-	91%	inferred high-confidence family alias from qwen3-30b-a3b (score=0.4819; benches=8)	Yes	Source
Qwen	Qwen 3 32B	-	93.80%	-	Yes	Source
Qwen	Qwen 3 235B A22B	-	95.60%	-	Yes	Source