Qwen3.5 35B A3B

Name: Qwen3.5 35B A3B
Author: Alibaba

Alibaba

Qwen3.5 35B A3B is a lightweight MoE model from the Qwen3.5 series with 3 billion active parameters. Fast and efficient with reasoning support.

Key Specifications

Parameters

35.0B

Context

Release Date

March 1, 2026

Average Score

69.0%

API Documentation Repository Model Weights Results Blog

Timeline

Key dates in the model's history

Announcement

March 1, 2026

Last Update

March 21, 2026

Today

September 10, 2026

Technical Specifications

Parameters

35.0B

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

69.2%

Reasoning

Logical reasoning and analysis

GPQA

Self-reported by the model provider • Self-reported

84.2%

Multimodal

Working with images and visual data

AI2D

TEST • Self-reported

93.0%

MMMU

Self-reported by the model provider • Self-reported

81.4%

Other Tests

Specialized benchmarks

CountBench

• Self-reported

98.0%

VLMsAreBlind

• Self-reported

97.0%

MMLU-Redux

• Self-reported

93.0%

With CI (best from options) • Self-reported

93.0%

IFEval

• Self-reported

92.0%

AA-LCR

Self-reported by the model provider • Self-reported

58.5%

AndroidWorld_SR

Self-reported by the model provider • Self-reported

71.1%

BabyVision

With CI (highest reported variant) • Self-reported

38.4%

BFCL-V4

Self-reported by the model provider • Self-reported

67.3%

BrowseComp

Self-reported by the model provider • Self-reported

61.0%

BrowseComp-zh

Self-reported by the model provider • Self-reported

69.5%

CC-OCR

Self-reported by the model provider • Self-reported

80.7%

C-Eval

Self-reported by the model provider • Self-reported

90.2%

CharXiv-R

RQ • Self-reported

77.5%

CodeForces

Elo Rating • Self-reported

82.2%

DeepPlanning

Self-reported by the model provider • Self-reported

22.8%

DynaMath

Self-reported by the model provider • Self-reported

85.0%

EmbSpatialBench

Self-reported by the model provider • Self-reported

83.1%

ERQA

Self-reported by the model provider • Self-reported

64.8%

FullStackBench en

Self-reported by the model provider • Self-reported

58.1%

FullStackBench zh

Self-reported by the model provider • Self-reported

55.0%

Global PIQA

Self-reported by the model provider • Self-reported

86.6%

Hallusion Bench

Self-reported by the model provider • Self-reported

67.9%

HMMT 2025

February 2025 • Self-reported

89.0%

HMMT25

November 2025 • Self-reported

89.2%

Humanity's Last Exam

With tools • Self-reported

47.4%

Hypersim

Self-reported by the model provider • Self-reported

13.1%

IFBench

Self-reported by the model provider • Self-reported

70.2%

Include

Self-reported by the model provider • Self-reported

79.7%

LingoQA

Self-reported by the model provider • Self-reported

79.2%

LiveCodeBench v6

Self-reported by the model provider • Self-reported

74.6%

LongBench v2

Self-reported by the model provider • Self-reported

59.0%

LVBench

Self-reported by the model provider • Self-reported

71.4%

MathVision

Self-reported by the model provider • Self-reported

83.9%

MathVista-Mini

Self-reported by the model provider • Self-reported

86.2%

MAXIFE

Self-reported by the model provider • Self-reported

86.6%

MedXpertQA

MM • Self-reported

61.4%

MLVU

Self-reported by the model provider • Self-reported

85.6%

MMBench-V1.1

EN_V1.1_dev • Self-reported

91.5%

MMLongBench-Doc

Self-reported by the model provider • Self-reported

59.5%

MMLU-Pro

Self-reported by the model provider • Self-reported

85.3%

MMLU-ProX

Self-reported by the model provider • Self-reported

81.0%

MMMLU

Self-reported by the model provider • Self-reported

85.2%

MMMU-Pro

Self-reported by the model provider • Self-reported

75.1%

MMStar

Self-reported by the model provider • Self-reported

81.9%

MMVU

Self-reported by the model provider • Self-reported

72.3%

Multi-Challenge

Self-reported by the model provider • Self-reported

60.0%

MVBench

Self-reported by the model provider • Self-reported

74.8%

NOVA-63

Self-reported by the model provider • Self-reported

57.1%

Nuscene

Self-reported by the model provider • Self-reported

14.6%

OCRBench

Self-reported by the model provider • Self-reported

91.0%

ODinW

13 • Self-reported

42.6%

OJBench

Self-reported by the model provider • Self-reported

36.0%

OmniDocBench 1.5

Self-reported by the model provider • Self-reported

89.3%

OSWorld-Verified

Self-reported by the model provider • Self-reported

54.5%

PMC-VQA

Self-reported by the model provider • Self-reported

62.0%

PolyMATH

Self-reported by the model provider • Self-reported

64.4%

RealWorldQA

Self-reported by the model provider • Self-reported

84.1%

RefCOCO-avg

Self-reported by the model provider • Self-reported

89.2%

RefSpatialBench

Self-reported by the model provider • Self-reported

63.5%

ScreenSpot Pro

Self-reported by the model provider • Self-reported

68.6%

Seal-0

Self-reported by the model provider • Self-reported

41.4%

SimpleVQA

Self-reported by the model provider • Self-reported

58.3%

SlakeVQA

Self-reported by the model provider • Self-reported

78.7%

SUNRGBD

Self-reported by the model provider • Self-reported

33.4%

SuperGPQA

Self-reported by the model provider • Self-reported

63.4%

t2-bench

Self-reported by the model provider • Self-reported

81.2%

Terminal-Bench 2.0

Self-reported by the model provider • Self-reported

40.5%

TIR-Bench

With CI (higher of reported variant) • Self-reported

55.5%

VideoMME w/o sub.

Without subtitles • Self-reported

82.5%

VideoMME w sub.

With subtitles • Self-reported

86.6%

VideoMMMU

Self-reported by the model provider • Self-reported

80.4%

VITA-Bench

Self-reported by the model provider • Self-reported

31.9%

WideSearch

Self-reported by the model provider • Self-reported

57.1%

WMT24++

Self-reported by the model provider • Self-reported

76.3%

ZEROBench

Self-reported by the model provider • Self-reported

8.0%

ZEROBench-Sub

Self-reported by the model provider • Self-reported

34.1%

License & Metadata

License

apache-2.0

Announcement Date

March 1, 2026

Last Updated

March 21, 2026

Articles about Qwen3.5 35B A3B

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

Qwen3.5 27B

Alibaba

27.0B

Best score:0.9 (GPQA)

Released:Mar 2026

Qwen2.5-Coder 32B Instruct

Alibaba

32.0B

Best score:0.9 (HumanEval)

Released:Sep 2024

Price:$0.09/1M tokens

Qwen3.5 9B

Alibaba

9.0B

Best score:0.8 (GPQA)

Released:Mar 2026

Qwen3 32B

Alibaba

32.8B

Released:Apr 2025

Price:$0.40/1M tokens

Qwen3-Next-80B-A3B-Instruct

Alibaba

80.0B

Best score:0.7 (GPQA)

Released:Sep 2025

Price:$0.15/1M tokens

Qwen2.5 14B Instruct

Alibaba

14.7B

Best score:0.8 (HumanEval)

Released:Sep 2024

QwQ-32B-Preview

Alibaba

32.5B

Best score:0.7 (GPQA)

Released:Nov 2024

Price:$1.20/1M tokens

Qwen3.7 Max

Alibaba

Best score:0.9 (GPQA)

Released:May 2026

Price:$2.50/1M tokens

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.