Atlas

Models

← All models

Qwen2.5 72B Instruct

Alibaba · 2024-09-19 · 72.7B parameters

Qwen2.5-72B-Instruct is Alibaba's instruction-tuned dense 72B-class language model, post-trained for chat, instruction following, coding, mathematics, structured data, and tool calling. Its final pre-training stage and shipped configuration use a 32,768-token sequence limit, while Qwen documents extrapolated operation to 131,072 tokens with YaRN and Dual Chunk Attention.

Benchmark scores

BenchmarkScore
AA-LCR20.3
AA-Omniscience Index-52.2
AIME 202514.0
Artificial Analysis Intelligence Index9.6
Artificial Analysis Omniscience Accuracy17.6
Artificial Analysis Omniscience Hallucination Rate84.7
BALROG16.2
BuzzBench45.9
Capability124.6
Confabulations Leaderboard Confabulation Rate32.2
Confabulations Leaderboard Non-Response Rate6.0
Confabulations Leaderboard Weighted Score19.1
Creative Writing v272.2
CritPt0.0
EDIT-Bench Complete45.2
EDIT-Bench Complete Easy73.2
EDIT-Bench Complete Hard11.4
EDIT-Bench Core53.7
EDIT-Bench Core Easy80.7
EDIT-Bench Core Hard17.4
EQ-Bench 3846
EQ-Bench 3 (Rubric Score)33.5
EQ-Bench 3: Analytic6.0
EQ-Bench 3: Assertive3.0
EQ-Bench 3: Compliant7.5
EQ-Bench 3: Empathy4.5
EQ-Bench 3: Humanlike2.8
EQ-Bench 3: Insight3.5
EQ-Bench 3: Moralising3.9
EQ-Bench 3: Pragmatic3.7
Loading Atlas data…