Qwen3.5 397B-A17B is Alibaba's flagship open-weight language model, featuring 397 billion total parameters with 512 Mixture-of-Experts activating just 17 billion per token. Released in February 2026, it supports 201 languages and a 256K-token context window, with multi-token prediction for accelerated inference. The model delivers state-of-the-art performance across language understanding, logical reasoning, code generation, and GUI interactions, while being 60% cheaper to run than its predecessor and 8× more capable at handling large concurrent workloads.
API|VisionReasoning|Open ModelApache 2.0
Knowledge Cutoff
2026-02-01
Input → Output Format
Context Memory
262KIN66KOUT
AI Performance Evaluation
Arena Overall Score
1446
±5As of 2026-04-23
Overall Rank
No.42
19,254 Votes
Arena by Ability
Hard Prompts
1463±6No.45
Expert Knowledge
1468±15No.41
Instruction Following
1428±8No.46
Conversation Memory
1450±11No.43
Creative
1415±11No.44
Coding
1487±9No.45
Math
1445±16No.36
Arena by Occupation
Creative Writing
1427±9No.38
Social Sciences
1463±11No.41
Media
1404±10No.51
Business
1441±10No.46
Healthcare
1466±16No.45
Legal
1439±16No.61
Software
1481±7No.41
Mathematics
1454±18No.32
Source:Arena Intelligence
Overall
AA Intelligence Index
40%↑2%
Reasoning & Math
GPQA Diamond
86%↑5%
HLE
19%↑2%
Coding
AA Coding Index
37%↑3%
TAU2
84%↑11%
TerminalBench
36%↑5%
SciCode
41%↑0%
Language & Instructions
IFBench
52%↓5%
AA-LCR
58%↓4%
Output Speed
Standard Mode
52tok/s↓30
First Output 1.51s
Reasoning Mode
62tok/s↓26
First Output 52.44s
Source:Artificial Analysis
