Atlas

Models

← All models

Qwen2-72B

Alibaba · 2024-06-07 · 72.7B parameters

Qwen2-72B is Alibaba's 72.7-billion-parameter base language model from the Qwen2 series, released June 7, 2024. It is a dense decoder-only Transformer with RMSNorm, SwiGLU, grouped-query attention, and RoPE, pretrained on a multilingual corpus of over 7 trillion tokens with increased code and mathematics coverage. Its concluding pretraining phase used 32,768-token sequences, while YaRN and Dual Chunk Attention enable extrapolated processing up to 131,072 tokens; as a base model it is intended for further post-training and underlies Qwen2-72B-Instruct.

Benchmark scores

BenchmarkScore
BBH (Open LLM Leaderboard v2)66.2
Capability128.8
Epoch Capabilities Index (ECI)125
GPQA (Open LLM Leaderboard v2)39.4
IFEval (Open LLM Leaderboard v2)38.2
MATH Level 5 (Open LLM Leaderboard v2)31.1
MMLU-Pro (Open LLM Leaderboard v2)57.3
MuSR (Open LLM Leaderboard v2)47.0
Loading Atlas data…