Atlas

Models

← All models

Qwen2.5 Instruct 32B

Alibaba · 2024-09-19 · 32.8B parameters

Qwen2.5-32B-Instruct is Alibaba's 32-billion-parameter instruction-tuned language model from the Qwen2.5 series, released September 19, 2024. It is a dense Transformer decoder with 128K-token context support (via YaRN extrapolation) and improvements over Qwen2 in instruction following, long-text generation, structured data understanding, and JSON output. The model is multilingual, covering over 29 languages, and shows strong performance in both coding and mathematics tasks.

Benchmark scores

BenchmarkScore
Artificial Analysis Intelligence Index7.5
BBH (Open LLM Leaderboard v2)69.1
BigCodeBench Complete52.3
BigCodeBench Instruct45.0
BigCodeBench-Hard Complete27.0
BigCodeBench-Hard Instruct22.3
BRIDGE Medical (chain-of-thought)37.7
BRIDGE Medical (few-shot)49.5
BRIDGE Medical (zero-shot)39.9
Capability124.7
EQ-Bench v279.9
EQ-Bench v2 + MAGI-Hard Combined78.3
Global PIQA - Non-Parallel (Strict Exact Match)75.2
Global PIQA - Parallel (Strict Exact Match)56.4
GPQA (Open LLM Leaderboard v2)33.8
Humanity's Last Exam (Text-Only)3.8
IFEval (Open LLM Leaderboard v2)83.5
LiveCodeBench (Artificial Analysis)24.8
MAGI-Hard76.7
MATH Level 556.1
MATH Level 5 (Open LLM Leaderboard v2)62.5
MATH-500 (Artificial Analysis source)80.5
MMLU Pro69.7
MMLU-Pro (Open LLM Leaderboard v2)56.7
MuSR (Open LLM Leaderboard v2)42.6
OTIS Mock AIME 2024-20257.4
SciCode (Artificial Analysis)22.9
ZeroEval CRUX69.9
ZeroEval MATH Level 559.5
ZeroEval MMLU-Redux83.7
Loading Atlas data…