Alibaba launches open source Qwen3 besting OpenAI o1

Chinese e-commerce and web giant Alibaba’s Qwen Team has officially started Open Source AI AI Sensol AI SERSE AI Great Language Multimodal models, visible between the most modern models for open models and approaching the property models, Openai and Google’s likes.

There are two “mixed expert” models and six intense models and six “mixed expert” models and six “mixed expert” models and six intense models for the QWen3 series. “Mixed Specialists” approach covers a combination of model types, combining models in the model’s internal parameters (known as parameters), a few different models activated. It was Open source is popular by French AI Startup Mistral.

According to the team, the 235 billion parameter version of GWEN3, Open Source of A22B, a open source of A22B open source R1 and Openai owners (500 user questions in software engineering and mathematics) and new, owned Google Gemini 2.5-Pro.

In general, Gwen3-235B-A22B, as one of the most powerful public models, as one of the most powerful public models, to achieve parity or advantage than large industrial proposals.

Hybrid (justification) theory

Gwen3 models allow users to pass users, science, mathematics, engineering and other specialized areas of “dynamic justification” or “dynamic justification”, and more consuming and more time and more time consuming and comprehensive substantial steps. This A approach developed by nous research and other AI Starting and research teams.

With Qwen3, users can attract more intensive “thinking mode” using the key mentioned in Qwen Chat website or including special tips /think or /no_think When placing the model via local or API, it allows you to use the task complexity depending on the complexity.

Users can now include and place these models in platforms as modellcope, Kaggle and Github, as well as direct contact with dozens of directories Qwen Chat web interface and mobile apps. The release includes a mixture of specialists (MOE) and dense models available under the Apache 2.0 open source license.

To date, when I use the Qwen Chat website, it was able to create an image by adherence to a relatively fast and worthy urgent – especially when it enters the image of the image. However, he wanted to enter the frequent and to the ordinary Chinese composition restrictions (such as forbidding or responding answers to Tiananmen Square protests).

In addition to Moe offers, Qwen3 includes dense models in different scales: Qwen3-32B, Qwen3-4B, Qwen3-4B, Qwen3-4B, Qwen3-4B, Qwen3-4b, Qwen3-4b, Qwen3-1.7b and Qwen3-1.7b and Qwen3-0.6b.

These models range in size and architecture, offers options to adapt to different needs and calculation budgets.

Qwen3 models are now significantly expanding multilingual support, now covers 119 languages ​​and dialects along the main language families. This expands potential applications of models that facilitate research and deployment in the context of extensive language.

Model Training and Architecture

In terms of model training, Qwen3 represents a big step from its predecessor, qwen2.5. Pretraingare database size has doubled up to 36 trillion token.

Data sources include document extract and synthetic content, such as PDF, generated using previous QWEN models of mathematics and coding.

The training pipeline consisted of three-stage post-training elegance to activate hybrid thinking and thinking. Training improvement allows you to adapt or put the performance of GWEN3 to the intense main models of GWEN3.

Placement options are versatile. Users can combine Qwen3 models using frames such as SGLANG and VLLM, both of which offer Openai matching points.

Options such as Local Use, LMStudio, MLX, Llama.CPP and Ktransformers are recommended. In addition, users interested in the capabilities of the models are encouraged to investigate the Qwen-Agent Toolbar, which facilitates the tool call operations.

Junyang Lin, Qwen Team member, commented in x This setup is critical, but the building is critical, but a building that solves more less glamorous technical problems, as a critical, but redempted multilingual performance without quality victims.

The Lin also indicated the team’s training agents focused on training agents, which are long horizontal for real world assignments.

What does the enterprise mean for decision makers

Engineering groups can display existing Openai matching points with a new model instead of weeks. Moe checkpoints (22 B are active 235 B settings and 3 B assets 30 B) GPT-4 class justification 20-30 b intense model’s GPU memory price.

Official LORA and QLORA hooks allow special delicate adjustment to the third party seller without sending property information.

Dense options from 0.6 B to 32 facilitate the expansion of prototype and modesty in laptops to many-GPU groups.

The weight of weights in the hotta can be entered and checked and checked out. Moe Sparlity reduces the number of active parameters per call by cutting the inferential attack surface.

Although the use of the Apache-2.0 license removes based legal barriers, organizations should review the exports and management results using a Chinese-based model.

Again, at the same time, at the same time DeepSeek, Tencent and Byrettion, as well as the above-mentioned Openai, Google, Microsoft, Anthropic, Amazon, Meta and others offering a countless and growing innocent alternative. Permissive Apache 2.0 License – Allowing unlimited commercial use – also a great advantage over other open source players such as more limiting meta.

In addition, the most powerful and accessible models between AI providers should try to be flexible and open to evaluate new models for essentials, AI agents and work flows, which want to reduce costs.

Looking forward

Qwen team increases QWen3 positions

For the next stage of Qwen, plans include expanding data and model size, expand the context length, increase the use of the context length, increase the use of modality with environmental feedback.

As the landscape of large-scale AI research continues to develop, an important stage from the open weight under a license of Gwen3, reduces obstacles in order to innovate with researchers, developers and modern LLMS.