Atlas

Models

← All models

Qwen1.5 MoE A2.7B Chat

Alibaba · 2024-03-28 · 14.3B parameters

Qwen1.5-MoE-A2.7B-Chat is Alibaba's chat-tuned Mixture-of-Experts model from the Qwen1.5 family, released March 28, 2024. The model has 14.3 billion total parameters and activates about 2.7 billion per token through 64 experts (4 shared plus 60 routed, with 4 routed experts selected), and Alibaba measured approximately 1.74× the inference speed of Qwen1.5-7B while reporting comparable performance. It was upcycled from the Qwen-1.8B dense model and supports a 32K-token context window.

Benchmark scores

BenchmarkScore
BBH (Open LLM Leaderboard v2)42.7
Capability101.3
EQ-Bench v258.1
EQ-Bench v2 + MAGI-Hard Combined48.2
GPQA (Open LLM Leaderboard v2)27.4
IFEval (Open LLM Leaderboard v2)38.0
MAGI-Hard38.3
MATH Level 5 (Open LLM Leaderboard v2)6.3
MMLU-Pro (Open LLM Leaderboard v2)29.2
MuSR (Open LLM Leaderboard v2)39.0
Loading Atlas data…