Knowledge benchmark

ARC (AI2) Challenge leaderboard

As of October 2026, DeepSeek-V3 has the highest published ARC (AI2) Challenge score on Noometry at 95.3%, out of 39 models with results.

Last verified October 10, 2026

About ARC (AI2) Challenge

Grade-school science multiple-choice questions that retrieval alone cannot answer.

Category: Knowledge
Introduced: 2018
Format: Multiple choice
Unit: Percent (random guessing ≈ 25%)
Official site: allenai.org

Top 15 models

DeepSeek-V3 95.3%
Llama 3.1-405B 95.3%
Qwen2.5 72B Instruct 94.5%
DeepSeek-V2 (MoE-236B, May 2024) 92.2%
phi-3-medium 14B 91.6%
Phi 3 Small 8k Instruct 90.7%
GPT-3.5-turbo 87.4%
Mixtral 8x7B 87.3%
Claude Instant 86.3%
Phi 3 Mini 4k Instruct 84.9%
Qwen-14B 84.4%
Llama 3-8B 82.8%
Mistral 7B 78.6%
Gemma 7B 78.3%
Llama 2-70B 78.3%
7580859095100

Sponsored placements are available on pages like this one. Advertise on Noometry

All results

ARC (AI2) Challenge results by model
#	Model	Provider	Score	Setting	Source
1	DeepSeek-V3	DeepSeek	95.3%		Epoch AI
2	Llama 3.1-405B	Meta	95.3%		Epoch AI
3	Qwen2.5 72B Instruct	Alibaba (Qwen)	94.5%		Epoch AI
4	DeepSeek-V2 (MoE-236B, May 2024)	DeepSeek	92.2%		Epoch AI
5	phi-3-medium 14B	Microsoft	91.6%		Epoch AI
6	Phi 3 Small 8k Instruct	Microsoft	90.7%		Epoch AI
7	GPT-3.5-turbo	OpenAI	87.4%		Epoch AI
8	Mixtral 8x7B	Mistral AI	87.3%		Epoch AI
9	Claude Instant	Anthropic	86.3%		Epoch AI
10	Phi 3 Mini 4k Instruct	Microsoft	84.9%		Epoch AI
11	Qwen-14B	Alibaba (Qwen)	84.4%		Epoch AI
12	Llama 3-8B	Meta	82.8%		Epoch AI
13	Mistral 7B	Mistral AI	78.6%		Epoch AI
14	Gemma 7B	Google	78.3%		Epoch AI
15	Llama 2-70B	Meta	78.3%		Epoch AI
16	Phi-2	Microsoft	75.9%		Epoch AI
17	Qwen-7B	Alibaba (Qwen)	75.3%		Epoch AI
18	Qwen2.5-Coder-32B	Alibaba (Qwen)	70.5%		Epoch AI
19	Falcon-180B	Technology Innovation Institute	67.8%		Epoch AI
20	Deepseek Coder v2	DeepSeek	64.3%		Epoch AI
21	Falcon-40B	Technology Innovation Institute	61.9%		Epoch AI
22	Qwen2.5-Coder (1.5B)	Alibaba (Qwen)	60.9%		Epoch AI
23	Llama 2-13B	Meta	60.3%		Epoch AI
24	Nemotron-4 15B	NVIDIA	55.5%		Epoch AI
25	INTELLECT-1	Hugging Face	54.5%		Epoch AI
26	Llama 2-34B	Meta	54.5%		Epoch AI
27	Llama 13b	Meta	52.7%		Epoch AI
28	Yi 6B	01.AI	50.3%		Epoch AI
29	Falcon-7B	Technology Innovation Institute	47.9%		Epoch AI
30	StarCoder 2 15B	NVIDIA	47.2%		Epoch AI
31	Llama 2-7B	Meta	45.9%		Epoch AI
32	Phi-1.5	Microsoft	44.4%	5	Epoch AI
33	DeepSeek Coder 33B	DeepSeek	42.2%		Epoch AI
34	Gemma 2B	Google	42.1%		Epoch AI
35	Dolly 2.0-12b	Databricks	39.6%		Epoch AI
36	StarCoder 2 7B	NVIDIA	38.7%		Epoch AI
37	DeepSeek Coder 6.7B	DeepSeek	36.4%		Epoch AI
38	StarCoder 2 3B	NVIDIA	34.2%		Epoch AI
39	DeepSeek Coder 1.3B	DeepSeek	25.4%		Epoch AI

Compare the leaders

Other knowledge benchmarks

Frequently asked questions

What does ARC (AI2) Challenge measure?

Grade-school science multiple-choice questions that retrieval alone cannot answer.

Which model has the highest ARC (AI2) Challenge score?

As of October 2026, DeepSeek-V3 has the highest published ARC (AI2) Challenge score on Noometry at 95.3%, out of 39 models with results.

What is the best open-weight model on ARC (AI2) Challenge?

DeepSeek-V3 has the highest ARC (AI2) Challenge accuracy among open-weight models at 95.3%, ranking 1 of 39 overall.