Skip to main content
Language modelActive

Qwen 3 VL 30B A3B Thinking

Alibaba

Released
September 30, 2025
Data date
September 8, 2026
Benchmark-Meta-Score
Unavailable
Fewer than eight test series
Speed
110.8 tokens/s
Output throughput
Price
$0.29 · $1
Input · Output / 1M tokens
Input
Text, Image, Video
Output
Text

Qwen 3 VL 30B A3B Thinking combines image and video understanding with a deliberately trained thinking mode. It accepts text, images, and video and can reason about visual information across space and time. Its Mixture-of-Experts architecture activates about 3.3 billion of 30.5 billion total parameters per token. That lowers active compute, but it does not make the model a 3.3B model.

The checkpoint provides 262,144 context tokens. Compared with the dense Qwen 3 VL 32B and the smaller Qwen 3 VL 8B Thinking, its defining trade-off is the balance between MoE capacity and active compute.

Specifications and access

SpecificationValue and source
Weights
Open weightsSource
Access
API, LocalSource
Parameters
30.5BSource
Active parameters
3.3BSource
Context window
262,144 tokensSource
Weights repository
Qwen/Qwen3-VL-30B-A3B-ThinkingSource
License
Apache-2.0Source
Architecture
vision encoder + MLP vision-language merger + Qwen3 language model; dense or MoESource
Input
Text, Image, VideoSource
Output
TextSource
Reasoning mode
thinkingSource
Tool calling
YesSource
Structured output
YesSource
Native context window
262,144 tokensSource
Maximum context window
262,144 tokensSource
Parameters
30.5 BSource
Active parameters
3.3 BSource
Languages
39 languages testedSource
Repository created
2025-09-30 (Repository creation date is a precise publication proxy, but may differ from the public announcement date or an individual checkpoint's first availability.)Source