Terminal-Bench 2.1

A verified refresh of Terminal-Bench 2.0: 89 curated tasks across software engineering, system administration, data processing, model training, and security, with environment and instruction fixes so scores reflect agent capability rather than environment gaps.

Models: 114 current, 243 in total
Median: 38.6%
Range: 0.0%–89.9%

Score distribution

Lowest: 0.0%
Median: 38.6%
Highest: 89.9%

Leaderboard

Include deprecated models

Rank	Model	Score
101	Ministral 3 8B Mistral	4.1%
102	Llama 4 Scout Meta	3.7%
103	NVIDIA Nemotron 3 Nano 4B NVIDIA	3.7%
104	Qwen3.5 2B (Reasoning) Alibaba	3.0%
105	Gemma 4 E4B (Reasoning) Google	1.9%
106	K2 Horizon 0.9B Institute of Foundation Models	1.1%
107	Nanbeige4.1-3B Nanbeige	1.1%
108	Gemma 4 E2B (Reasoning) Google	0.4%
109	Phi-4 Mini Instruct Microsoft	0.4%
110	Qwen3.5 0.8B (Non-reasoning) Alibaba	0.4%
111	MiniCPM-V 4.6 1.3B OpenBMB	0.0%
112	Ministral 3 3B Mistral	0.0%
113	Qwen3.5 0.8B (Reasoning) Alibaba	0.0%
114	Qwen3.5 2B (Non-reasoning) Alibaba	0.0%

101
Ministral 3 8B
Mistral
4.1%
102
Llama 4 Scout

Data from Artificial Analysis

Command Palette

Terminal-Bench 2.1

Score distribution

Leaderboard