跳转到主要内容

Qwen3 Coder 480B A35B Instruct

AlibabaQwenOpen WeightApache 2.0 · Commercial OK

描述

Qwen3-Coder-480B-A35B-Instruct is Qwen's most agentic code model to date, featuring 480 billion total parameters with 35 billion activated parameters using MoE architecture. It achieves significant performance among open models on Agentic Coding, Agentic Browser-Use, and foundational coding tasks, with results comparable to Claude Sonnet. Features native 256K token context length (extendable to 1M tokens with Yarn), optimized for repository-scale understanding, and specialized function call format for agentic coding platforms like Qwen Code and CLINE.

发布日期
2025-07-22
参数规模
480.0B
上下文长度
262K
支持模态
text

能力雷达图

38
general
38
coding
52
reasoning
41
science估算
60
agents
0
multimodal

Science 在缺少专门科学评测时使用推理能力代理估算。

排行榜排名

领域#排名分数来源
智能体与工具55
56.0
LS
代码能力榜170
47.0
AA
通用能力榜211
49.0
AA
数学推理170
53.0
AA
科学能力249
43.0
AA

基准测试分数 (LLM Stats)

Agents

BFCL-v368.7%自报
MM-Mind2Web55.8%自报
Terminal-Bench 2.037.5%自报

Code

SWE-Bench Verified69.6%自报
Aider-Polyglot61.8%自报
SWE-bench Multilingual54.7%自报
Multi-SWE-Bench25.8%自报

Communication

TAU-bench Retail77.5%自报
TAU-bench Airline60.0%自报

Language

Spider31.1%自报

AA 评测指数

Math Index
39.3
Intelligence Index
24.8
Coding Index
24.6
Math 500
0.9
Mmlu Pro
0.8
Gpqa
0.6
Livecodebench
0.6
Aime
0.5
Tau2
0.4
Lcr
0.4
Ifbench
0.4
Aime 25
0.4
Scicode
0.4
Terminalbench Hard
0.2
Hle
0.0

LLM Stats 分类评分

Structured Output
70
Communication
70
Finance
70
General
70
Tool Calling
60
Frontend Development
60
Multimodal
60
Reasoning
60
Agents
50
Code
50

定价

输入价格$0.3 / 1M tokens
输出价格$1.8 / 1M tokens
混合价格(3:1)$0.675 / 1M tokens

速度

Tokens/秒66.1 tokens/s
首Token延迟1.63s
首回答延迟1.63s

可用提供商

(LS 内部计价单位)

暂无提供商数据

外部链接