Alibaba logo

Qwen3.5 27B

Alibaba

Qwen3.5 27B is a mid-size model from the Qwen3.5 series with extended reasoning support. Optimal balance of intelligence and computational requirements.

Key Specifications

Parameters
27.0B
Context
-
Release Date
March 1, 2026
Average Score
70.3%

Timeline

Key dates in the model's history
Announcement
March 1, 2026
Last Update
March 21, 2026
Today
September 10, 2026

Technical Specifications

Parameters
27.0B
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
72.4%

Reasoning

Logical reasoning and analysis
GPQA
Self-reported
85.5%

Multimodal

Working with images and visual data
AI2D
TESTSelf-reported
93.0%
MMMU
Self-reported
82.3%

Other Tests

Specialized benchmarks
CountBench
Self-reported
98.0%
VLMsAreBlind
Self-reported
97.0%
IFEval
Self-reported
95.0%
V*
With CI (best from options)Self-reported
94.0%
MMLU-Redux
Self-reported
93.0%
AA-LCR
Self-reported
66.1%
AndroidWorld_SR
Self-reported
64.2%
BabyVision
With CI (highest reported variant)Self-reported
44.6%
BFCL-V4
Self-reported
68.5%
BrowseComp
Self-reported
61.0%
BrowseComp-zh
Self-reported
62.1%
C-Eval
Self-reported
90.5%
CC-OCR
Self-reported
81.0%
CharXiv-R
RQSelf-reported
79.5%
CodeForces
Elo RatingSelf-reported
80.7%
DeepPlanning
Self-reported
22.6%
DynaMath
Self-reported
87.7%
EmbSpatialBench
Self-reported
84.5%
ERQA
Self-reported
60.5%
FullStackBench en
Self-reported
60.1%
FullStackBench zh
Self-reported
57.4%
Global PIQA
Self-reported
87.5%
Hallusion Bench
Self-reported
70.0%
HMMT 2025
February 2025Self-reported
92.0%
HMMT25
November 2025Self-reported
89.8%
Humanity's Last Exam
With toolsSelf-reported
48.5%
Hypersim
Self-reported
13.0%
IFBench
Self-reported
76.5%
Include
Self-reported
81.6%
LingoQA
Self-reported
82.0%
LiveCodeBench v6
Self-reported
80.7%
LongBench v2
Self-reported
60.6%
LVBench
Self-reported
73.6%
MathVision
Self-reported
86.0%
MathVista-Mini
Self-reported
87.8%
MAXIFE
Self-reported
88.0%
MedXpertQA
MMSelf-reported
62.4%
MLVU
Self-reported
85.9%
MMBench-V1.1
EN_V1.1_devSelf-reported
92.6%
MMLongBench-Doc
Self-reported
60.2%
MMLU-Pro
Self-reported
86.1%
MMLU-ProX
Self-reported
82.2%
MMMLU
Self-reported
85.9%
MMMU-Pro
Self-reported
75.0%
MMStar
Self-reported
81.0%
MMVU
Self-reported
73.3%
Multi-Challenge
Self-reported
60.8%
MVBench
Self-reported
74.6%
NOVA-63
Self-reported
58.1%
Nuscene
Self-reported
15.2%
OCRBench
Self-reported
89.4%
ODinW
13Self-reported
41.1%
OJBench
Self-reported
40.1%
OmniDocBench 1.5
Self-reported
88.9%
OSWorld-Verified
Self-reported
56.2%
PMC-VQA
Self-reported
62.4%
PolyMATH
Self-reported
71.2%
RealWorldQA
Self-reported
83.7%
RefCOCO-avg
Self-reported
90.9%
RefSpatialBench
Self-reported
67.7%
ScreenSpot Pro
Self-reported
70.3%
Seal-0
Self-reported
47.2%
SimpleVQA
Self-reported
56.0%
SlakeVQA
Self-reported
80.0%
SUNRGBD
Self-reported
35.4%
SuperGPQA
Self-reported
65.6%
t2-bench
Self-reported
79.0%
Terminal-Bench 2.0
Self-reported
41.6%
TIR-Bench
With CI (higher of reported variant)Self-reported
59.8%
VideoMME w/o sub.
Without subtitlesSelf-reported
82.8%
VideoMME w sub.
With subtitlesSelf-reported
87.0%
VideoMMMU
Self-reported
82.3%
VITA-Bench
Self-reported
41.9%
WideSearch
Self-reported
61.1%
WMT24++
Self-reported
77.6%
ZEROBench
Self-reported
10.0%
ZEROBench-Sub
Self-reported
36.2%

License & Metadata

License
apache-2.0
Announcement Date
March 1, 2026
Last Updated
March 21, 2026

Articles about Qwen3.5 27B

A $600 Card That Runs 10,000 Tokens Per Second

A $600 Card That Runs 10,000 Tokens Per Second

Taalas wants to hardwire LLMs directly into silicon. Their ASIC approach delivers 10x Cerebras speed at a fraction of GPU power draw — but each chip runs exactly one model.

5 min
Intel's $949 GPU Has 32GB of VRAM. The Local AI Community Is Paying Attention.

Intel's $949 GPU Has 32GB of VRAM. The Local AI Community Is Paying Attention.

The Arc Pro B70 undercuts NVIDIA by half on price and beats it on VRAM. But Intel's software stack remains the elephant in the room.

6 min
The Best GPU for Local AI in 2026 Costs $650 — And It's from 2020

The Best GPU for Local AI in 2026 Costs $650 — And It's from 2020

Used RTX 3090 prices have cratered to $650 while RTX 5090s sell for $3,500. For the local LLM community, old hardware has never made more sense.

6 min
The Two Loops: How China's Open-Source AI Strategy Is Outpacing America

The Two Loops: How China's Open-Source AI Strategy Is Outpacing America

A new USCC report warns that China's open AI models now dominate global downloads. 80% of US startups use Chinese models. Washington is scrambling.

9 min
Unsloth Studio Wants to Be the IDE for Local AI — Training Included

Unsloth Studio Wants to Be the IDE for Local AI — Training Included

The open-source tool combines inference and fine-tuning in one interface, with 70% less VRAM and no-code training for 500+ models. LM Studio should be nervous.

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

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.

7 min

ik_llama.cpp Delivers 26x Faster Prompt Processing for Qwen 3.5

A new optimized C++ inference engine achieves 26x speedup on Qwen 3.5 prompt processing, a major win for local AI deployment.

2 min

Similar Models

All Models

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.