Meet Qwen 3: a new AI model that promises to think deeper and act faster
29 April 11:37
On April 29, Chinese tech company Alibaba unveiled a new line of Qwen3 language models. According to the developers, the new generation of artificial intelligence from Qwen is significantly improved compared to previous versions, "Komersant Ukrainian" reports.
What is special about Qwen3
The flagship model Qwen3-235B-A22B has an impressive 235 billion parameters, of which 22 billion are actively used. It is designed to successfully compete with such industry giants as DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro. Even the small Qwen3-4B model demonstrates results at the level of the previous generation Qwen2.5-72B-Instruct, which demonstrates the effectiveness of new approaches in development.
Two modes of thinking
The most interesting feature of the new line is the introduction of hybrid thinking modes:
- Thinking Mode – the model thinks through a task in detail step by step before giving an answer, which is ideal for complex problems.
- Non-Thinking Mode – provides near-instantanswers to simpler questions.
Users can switch between these modes using the /think and /no_think commands, controlling the balance between the speed and depth of the model’s thinking.
Multilingualism on a new scale
Qwen3 supports 119 languages and dialects, including Indo-European, Sino-Tibetan, Afro-Asian, and many other language families. This makes the model accessible to users from different parts of the world and opens up new opportunities for international application.
The AI also speaks Ukrainian.
Читайте нас у Telegram: головні новини коротко
A series of models for different needs
The developers have released a whole line of models of different sizes:
- Two MoE models: Qwen3-235B-A22B (the most powerful) and Qwen3-30B-A3B.
- Six dense models: from Qwen3-32B to Qwen3-0.6B.
All models are available with open scales under the Apache 2.0 license, which allows them to be used for both research and commercial projects.
Improved development and training
The process of creating Qwen3 included training on 36 trillion tokens – almost twice as many as in the previous generation. The training took place in three stages with gradual data complexity and an increase in the context window to 128 thousand tokens.
Powerful agent capabilities
The developers paid special attention to integration with tools and agent functions. The model works perfectly with the Qwen-Agent framework, which simplifies the creation of AI assistants capable of interacting with various services and performing complex tasks.
Future prospects
The Qwen team sees the release of Qwen3 as an important step towards the creation of artificial general intelligence (AGI) and superintelligent AI (ASI). In the future, they plan to improve model architecture, increase data volumes and model sizes, expand the context window, and develop reinforcement learning.
Qwen3 is already available for use via Qwen Chat Web(chat.qwen.ai) and mobile app, as well as on Hugging Face, ModelScope, and Kaggle.
This new series of models adds to the ecosystem of natural language tools and provides developers with additional opportunities to create a variety of applications.
Читайте нас у Telegram: головні новини коротко
