Alibaba releases Qwen3

In partnership with

In Today’s Issue (no fluff):

• Alibaba releases Qwen3

• OpenAI rushes to fix bug that allowed sexual chats with minors

Read time: 3 minutes

Alibaba on Monday unveiled Qwen3, a family of AI models the company says matches or beats leading systems from Google and OpenAI. The models, ranging from 0.6 billion to 235 billion parameters, are available under an open license on Hugging Face and GitHub.

Qwen3 models introduce "hybrid" reasoning modes, letting users toggle between fast responses and more careful, fact-checked answers. Some versions also use mixture-of-experts (MoE) architecture for efficiency. The largest public model, Qwen3-32B, outperforms OpenAI’s o1 on coding benchmarks, while Alibaba’s internal tests show the unreleased Qwen-3-235B rivaling OpenAI’s o3-mini and Google’s Gemini 2.5 Pro.

Qwen3 models use hybrid reasoning modes and, in some versions, a mixture-of-experts architecture for greater efficiency. Alibaba says the models support 119 languages and were trained on nearly 36 trillion tokens from textbooks, Q&A data, code, and synthetic sources.

Technical details:

• Model Range: 0.5B to 72B parameters (open models), 235B for internal use

• License: Open-source (most models) via Hugging Face and GitHub

• Training Data: ~35.9 trillion tokens from books, code, Q&A, and synthetic content

• Languages: Supports 119 languages (with 29 officially evaluated)

• Architecture: Transformer, mixture-of-experts in select models

• Performance: Qwen3-32B outperforms OpenAI o1-mini on coding benchmarks

• Variants: Base, chat, and instruction-tuned versions available