# Qwen Releases 'Thinking' Multimodal MoE Model

> The new 30-billion-parameter Mixture-of-Experts model from Alibaba's Qwen team is designed to show its reasoning process for complex multimodal tasks.

Published by The Open Weights on Sep 15, 2025. Canonical: https://theopenweights.com/news/qwen3-omni-30b-a3b-thinking-1f1p

## Key facts

- Company: Qwen · Alibaba
- Model: Qwen3-Omni-30B-A3B-Thinking
- Version: Qwen3-Omni Thinking
- Category: Any-to-Any
- Modalities: Any-to-Any, Reasoning, Vision-Language
- License: Other (Open weights, commercial use allowed)
- Parameters: 30B (mixture of experts)
- Active parameters: 3B
- Context window: 66K tokens
- Architecture: Qwen3OmniMoeForConditionalGeneration
- Precision: BF16
- Weights size: 63.4 GB
- Languages: 119 languages
- Significance: notable
- Published: 2025-09-15
- Last verified: 2026-06-17
- Hugging Face: https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking
- Canonical URL: https://theopenweights.com/news/qwen3-omni-30b-a3b-thinking-1f1p

Alibaba's Qwen team has introduced a new model designed for transparent reasoning, called Qwen3-Omni-30B-A3B-Thinking. This release is a specialized variant of the broader Qwen3-Omni family, focusing on tasks that require complex, multi-step logic across various data types.

The model's key feature is its ability to output its "chain of thought," a step-by-step trace of its reasoning process. This transparency is a significant advantage for developers, allowing them to better understand, debug, and guide the model's decision-making. Instead of just providing a final answer, Qwen3-Omni shows its work, demystifying the path it took to reach a conclusion.

## Efficient Architecture, Broad Capabilities

Under the hood, Qwen3-Omni is a Mixture-of-Experts (MoE) model. While it contains a total of 30 billion parameters, it only activates an average of 3 billion for any given task. This architecture aims to provide the knowledge scale of a large model with the inference efficiency closer to that of a much smaller one.

As an "omni-modal" model, its capabilities extend beyond text to a wide range of inputs:
* Image and video understanding
* Audio processing
* Document analysis

This versatility makes it suitable for complex applications that need to synthesize information from multiple sources. The model and its weights are available for developers to explore on [its Hugging Face repository](https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking). It's released under a custom license agreement, so users should review the terms before deployment.

## Get the model

- [Hugging Face](https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking)

## Sources

- [Qwen/Qwen3-Omni-30B-A3B-Thinking](https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking) — Hugging Face, Sep 15, 2025

---
Source: The Open Weights (https://theopenweights.com/). Aggregated and written by Claude, curated by humans. Cite as: "Qwen Releases 'Thinking' Multimodal MoE Model", The Open Weights, Sep 15, 2025, https://theopenweights.com/news/qwen3-omni-30b-a3b-thinking-1f1p