Atlas

Models

← All models

Qwen2.5 3B Instruct

Alibaba · 2024-09-19 · 3.1B parameters

Qwen2.5-3B-Instruct is Alibaba's 3-billion-parameter instruction-tuned small language model from the Qwen2.5 series, released September 19, 2024. It supports more than 29 languages and a 32K-token context window and targets reasoning, coding, multilingual, and structured-output tasks.

Benchmark scores

BenchmarkScore
BBH (Open LLM Leaderboard v2)46.9
BRIDGE Medical (chain-of-thought)25.4
BRIDGE Medical (few-shot)37.2
BRIDGE Medical (zero-shot)26.6
Capability104.3
EQ-Bench v249.8
EQ-Bench v2 + MAGI-Hard Combined49.3
Global PIQA - Non-Parallel (Strict Exact Match)60.2
Global PIQA - Parallel (Strict Exact Match)35.4
GPQA (Open LLM Leaderboard v2)27.3
IFEval (Open LLM Leaderboard v2)64.7
MAGI-Hard48.8
MATH Level 5 (Open LLM Leaderboard v2)36.8
MEDIC (clinical summarization)85.6
MEDIC (closed-ended)49.6
MEDIC (open-ended Elo)1249
MMLU-Pro (Open LLM Leaderboard v2)32.5
MuSR (Open LLM Leaderboard v2)39.7
ZeroEval31.9
ZeroEval CRUX33.1
ZeroEval MATH Level 525.5
ZeroEval MMLU-Redux64.3
ZeroEval ZebraLogic4.8
Loading Atlas data…