Qwen3-Coder-480B-A35B-Instruct

Name: Qwen3-Coder-480B-A35B-Instruct
Author: Alibaba

Alibaba · 2025-07-22 · 480.2B parameters

Qwen3-Coder-480B-A35B-Instruct is Alibaba's instruction-tuned, code-specialized open-weight MoE model, released July 22, 2025, and Qwen reports 480B total and 35B activated parameters across 160 experts. It targets agentic coding, including function calling, tool use, and repository-scale reasoning, and supports 262,144 tokens natively with YaRN extension up to 1,048,576 tokens. The model supports only non-thinking mode.

Benchmark scores

Benchmark	Score
AA-LCR	42.3
AA-LCR v1.1	45.7
AA-Omniscience Index	-22.2
AHK-Eval	50.0
AHK-Eval Algorithms	50.0
AHK-Eval Data Structures	33.3
AHK-Eval Date and Time	0.0
AHK-Eval Easy	58.3
AHK-Eval Hard	33.3
AHK-Eval Hidden Cases	47.5
AHK-Eval Mid	41.7
AHK-Eval Numbers	83.3
AHK-Eval Parse Success	88.9
AHK-Eval Regex	50.0
AHK-Eval Strings	50.0
AIME 2025 (Artificial Analysis avg@10)	39.3
ALE-Bench	822
AlgoTune	1.4
Arena Code WebDev — Brand & Marketing	1289
Arena Code WebDev — Consumer Product	1275
Arena Code WebDev — Content Creation Tools	1263
Arena Code WebDev — Data & Analytics	1274
Arena Code WebDev — Frontend	1273
Arena Code WebDev — Gaming	1259
Arena Code WebDev — HTML	1287
Arena Code WebDev — React	1257
Arena Code WebDev — Reference-Based Design	1264
Arena Code WebDev — Simulations	1266
Arena Text — Business, Management & Financial Ops (No Style Control)	1343
Arena Text — Chinese (No Style Control)	1357