Atlas

Models

← All models

Qwen3 Coder 30B A3B Instruct

Alibaba · 2025-07-31 · 30.5B parameters

Qwen3-Coder-30B-A3B-Instruct is Alibaba's open-weight, code-specialized mixture-of-experts instruction model for agentic coding, repository-scale understanding, and tool use. Its checkpoint contains exactly 30,532,122,624 parameters, while Qwen reports 3.3B activated only at rounded precision; it supports non-thinking operation only. The native context is 262,144 tokens, and the release advertises YaRN extension to 1M; the shipped tokenizer configuration sets an exact maximum of 1,048,576.

Benchmark scores

BenchmarkScore
AA-LCR29.0
AA-Omniscience Index-50.9
AIME 202529.0
Artificial Analysis Intelligence Index13.6
Artificial Analysis Omniscience Accuracy15.8
Artificial Analysis Omniscience Hallucination Rate79.2
Artificial Analysis Openness Index44.4
BuseyBench SVG2.4
Capability127.5
CritPt0.0
GPQA Diamond51.6
Humanity's Last Exam (Text-Only)4.0
IFBench (Artificial Analysis)32.7
LiveCodeBench (Artificial Analysis)40.3
MATH-500 (Artificial Analysis source)89.3
MMLU Pro70.6
ObviousBench46.5
SciCode (Artificial Analysis)27.8
SnakeBench Average Apples1.9
SnakeBench Best Apples9.0
SnakeBench Rating22.7
SnakeBench Total Apples74.0
SnakeBench Win Rate47.2
Terminal-Bench Hard15.2
τ²-bench (Telecom)34.5
Loading Atlas data…