Atlas

Models

← All models

Qwen3 4B

Alibaba · 2025-04-29 · 4B parameters

Qwen3-4B is Alibaba's dense 4B open-weight model from the initial Qwen3 release in April 2025. It supports hybrid thinking/non-thinking operation, with 32,768 tokens of native context and a validated maximum of 131,072 tokens using YaRN, under an Apache 2.0 license.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-LCR v1.10.0
AIME 2025 (Artificial Analysis avg@10)22.3
Artificial Analysis Progress Index v1182
BRIDGE Medical (chain-of-thought)37.0
BRIDGE Medical (few-shot)42.8
BRIDGE Medical (zero-shot)38.5
Capability122.6
Chess Puzzles1.0
Global PIQA - Non-Parallel (Strict Exact Match)70.6
Global PIQA - Parallel (Strict Exact Match)53.0
GPQA Diamond52.3
GPQA Diamond (Artificial Analysis avg@5)52.2
Humanity's Last Exam (Text-Only)4.4
IFBench (Artificial Analysis avg@5)32.5
LiveCodeBench (Artificial Analysis)46.5
Longform Creative Writing17.0
MATH-500 (Artificial Analysis source)93.3
MEDIC (clinical summarization)85.0
MEDIC (closed-ended)56.8
MEDIC (open-ended Elo)1653
MMLU Pro69.6
SciCode (Artificial Analysis avg@3)16.7
Slop Score62.2
SnakeBench Average Apples1.4
SnakeBench Best Apples5.0
SnakeBench Rating16.4
SnakeBench Total Apples19.0
SnakeBench Win Rate46.2
Vectara Hallucination Leaderboard5.7
Loading Atlas data…