Atlas

Models

← All models

Qwen3 4B Instruct 2507

Alibaba · 2025-08-06 · 4B parameters

Qwen3-4B-Instruct-2507 is Alibaba's non-thinking instruction-tuned update to the dense 4B Qwen3 model, released in August 2025. It has a native 262,144-token context; the current official tokenizer configuration sets a 1,010,000-token maximum for the Qwen3-2507 family's documented million-token extension. It is a separate checkpoint from Qwen3-4B-Thinking-2507.

Benchmark scores

BenchmarkScore
AA-LCR7.3
AA-Omniscience Index-59.6
AIME 202552.3
Artificial Analysis Intelligence Index7.1
Artificial Analysis Omniscience Accuracy10.3
Artificial Analysis Omniscience Hallucination Rate78.0
Artificial Analysis Openness Index44.4
BFCL v4 Overall Accuracy35.7
Capability125.8
CritPt0.0
GPQA Diamond51.7
Humanity's Last Exam (Text-Only)4.7
IFBench (Artificial Analysis)33.5
LiveCodeBench (Artificial Analysis)37.7
MMLU Pro67.2
SciCode (Artificial Analysis)18.1
Terminal-Bench Hard4.5
τ²-bench (Telecom)26.6
Loading Atlas data…