Atlas

Models

← All models

Qwen3-Coder-Next

Alibaba · 2026-02-03 · 79.7B parameters

Qwen3-Coder-Next is Alibaba's open-weight, non-thinking coding model built on the Qwen3-Next hybrid Gated DeltaNet/Gated Attention mixture-of-experts architecture. Qwen reports 80B total and 3B active parameters at rounded precision, while the released checkpoint contains exactly 79,674,391,296 parameters. It is trained for agentic coding and long-horizon tool use, supports 262,144 tokens natively, and can be extended to 1,048,576 tokens with YaRN.

Benchmark scores

BenchmarkScore
AA-LCR40.0
AA-Omniscience Index-60.7
Artificial Analysis Agentic Index8.8
Artificial Analysis Coding Index36.2
Artificial Analysis Intelligence Index21.1
Artificial Analysis Omniscience Accuracy15.8
Artificial Analysis Omniscience Hallucination Rate90.9
Artificial Analysis Openness Index41.7
BuseyBench SVG3.2
Capability138.5
CritPt0.0
GDPval-AA v2718
GDPval-AA v2 (normalized)10.9
Global MMLU Lite79.4
GPQA Diamond73.7
Humanity's Last Exam (Text-Only)9.3
IFBench (Artificial Analysis)35.2
ObviousBench43.1
Ruby LLM Benchmarks — Program Fixer56.7
Ruby LLM Program Fixer — Calendar69.5
Ruby LLM Program Fixer — Parking Garage59.1
Ruby LLM Program Fixer — School Library52.8
Ruby LLM Program Fixer — Vending Machine45.4
SciCode32.3
SciCode (Artificial Analysis)32.3
SnakeBench Average Apples5.5
SnakeBench Best Apples16.0
SnakeBench Rating20.7
SnakeBench Total Apples55.0
SnakeBench Win Rate60.0
Loading Atlas data…