Qwen3.5-397B-A17B

Name: Qwen3.5-397B-A17B
Author: Alibaba

Multimodal

Alibaba

Qwen3.5-397B-A17B is the flagship Qwen model with a Mixture-of-Experts architecture containing 397 billion parameters and 17 billion active. Delivers state-of-the-art performance in knowledge, reasoning, coding, math, multilingual understanding, instruction following, and agentic tasks.

Key Specifications

Parameters

397.0B

Context

Release Date

February 15, 2026

Average Score

73.1%

API Documentation Repository Model Weights Results Blog

Timeline

Key dates in the model's history

Announcement

February 15, 2026

Last Update

February 17, 2026

Today

September 10, 2026

Technical Specifications

Parameters

397.0B

Training Tokens

Knowledge Cutoff

Family

Capabilities

MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Programming

Programming skills tests

SWE-Bench Verified

Self-reported by the model provider • Self-reported

76.4%

Reasoning

Logical reasoning and analysis

GPQA

Self-reported by the model provider • Self-reported

88.4%

Other Tests

Specialized benchmarks

MMLU-Redux

• Self-reported

95.0%

HMMT 2025

February 2025 • Self-reported

95.0%

C-Eval

• Self-reported

93.0%

HMMT25

November 2025 • Self-reported

93.0%

IFEval

• Self-reported

93.0%

AIME 2026

• Self-reported

91.0%

AA-LCR

Self-reported by the model provider • Self-reported

68.7%

BFCL-V4

Self-reported by the model provider • Self-reported

72.9%

BrowseComp

Self-reported by the model provider • Self-reported

69.0%

BrowseComp-zh

Self-reported by the model provider • Self-reported

70.3%

DeepPlanning

Self-reported by the model provider • Self-reported

34.3%

Global PIQA

Self-reported by the model provider • Self-reported

89.8%

Humanity's Last Exam

Self-reported by the model provider • Self-reported

28.7%

IFBench

Self-reported by the model provider • Self-reported

76.5%

IMO-AnswerBench

Self-reported by the model provider • Self-reported

80.9%

Include

Self-reported by the model provider • Self-reported

85.6%

LiveCodeBench v6

Self-reported by the model provider • Self-reported

83.6%

LongBench v2

Self-reported by the model provider • Self-reported

63.2%

MAXIFE

Self-reported by the model provider • Self-reported

88.2%

MCP-Mark

Self-reported by the model provider • Self-reported

46.1%

MMLU-Pro

Self-reported by the model provider • Self-reported

87.8%

MMLU-ProX

Self-reported by the model provider • Self-reported

84.7%

MMMLU

Self-reported by the model provider • Self-reported

88.5%

Multi-Challenge

Self-reported by the model provider • Self-reported

67.6%

NOVA-63

Self-reported by the model provider • Self-reported

59.1%

PolyMATH

Self-reported by the model provider • Self-reported

73.3%

Seal-0

Self-reported by the model provider • Self-reported

46.9%

SecCodeBench

Self-reported by the model provider • Self-reported

68.3%

SuperGPQA

Self-reported by the model provider • Self-reported

70.4%

SWE-bench Multilingual

Self-reported by the model provider • Self-reported

69.3%

t2-bench

Self-reported by the model provider • Self-reported

86.7%

Terminal-Bench 2.0

Self-reported by the model provider • Self-reported

52.5%

Toolathlon

Self-reported by the model provider • Self-reported

38.3%

VITA-Bench

Self-reported by the model provider • Self-reported

49.7%

WideSearch

Self-reported by the model provider • Self-reported

74.0%

WMT24++

Self-reported by the model provider • Self-reported

78.9%

License & Metadata

License

apache-2.0

Announcement Date

February 15, 2026

Last Updated

February 17, 2026

Articles about Qwen3.5-397B-A17B

Qwen 3.6 Quietly Appears on OpenRouter With a Million-Token Context

Alibaba's Qwen 3.6 Plus Preview surfaces on OpenRouter with no announcement, bringing 1M context window, mandatory reasoning, and a hybrid MoE architecture.

April 1, 2026

3 min

Two DGX Sparks vs One Mac Studio: Running 397B at Home

A Reddit user tested Qwen3.5-397B on dual NVIDIA DGX Sparks and a Mac Studio M3 Ultra 512GB. The results split the community — neither machine wins outright.

March 28, 2026

6 min

Alibaba Doubles Down on Open Source at ModelScope DevCon

Alibaba confirmed its commitment to open-sourcing Qwen, Wan, and new model families at ModelScope DevCon in Nanjing. Here's what was announced.

March 24, 2026

2 min

Qwen 3.5 Is Alibaba's Bid to Win the Agentic AI Era

Alibaba's Qwen 3.5 family uses extreme MoE efficiency to beat models 7x its size. The flagship is now live on Arena — and claims to outperform GPT-5.2.

March 23, 2026

7 min

Similar Models

All Models

Qwen3.8 Max

Alibaba

MM2.4T

Best score:0.9 (GPQA)

Released:Aug 2026

Price:$2.50/1M tokens

Qwen3.8 Flash

Alibaba

MM125.0B

Best score:0.9 (GPQA)

Released:Aug 2026

Price:$0.15/1M tokens

Qwen3.8 Flash Next

Alibaba

MM180.0B

Best score:0.9 (GPQA)

Released:Aug 2026

Price:$0.16/1M tokens

Qwen3 VL 32B Thinking

Alibaba

MM33.0B

Best score:0.9 (MMLU)

Released:Sep 2025

Qwen3.6 Plus

Alibaba

Best score:0.9 (GPQA)

Released:Mar 2026

Price:$0.50/1M tokens

Qwen3.6-27B

Alibaba

MM27.8B

Best score:0.9 (GPQA)

Released:Apr 2026

Price:$0.60/1M tokens

Qwen3.7-Plus

Alibaba

Best score:0.9 (GPQA)

Released:May 2026

Price:$0.32/1M tokens

Qwen3.6-35B-A3B

Alibaba

MM35.0B

Best score:0.9 (GPQA)

Released:Apr 2026

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.