Qwen3 is putting serious pressure on OpenAI and Google

AI MODELS

Qwen3 is putting serious pressure on OpenAI and Google

Alibaba has just rolled out Qwen3, a new family of AI models that it says can match, and sometimes even outperform, top names like OpenAI and Google.

The Qwen3 line-up covers models from 0.6 billion to 235 billion parameters, with most set to be available for download under an open licence on Hugging Face and GitHub.

Generally, the more parameters a model has, the better it is at solving problems.

These models are built to be "hybrid," meaning they can quickly handle simple tasks or take a bit more time to reason through tougher ones.

Some versions also use a "mixture of experts" system, breaking jobs into smaller tasks handled by specialist models to boost efficiency.

A few important points:

• Trained on 36 trillion tokens from textbooks, Q&A sets, code, and AI-generated content.

• Supports 119 languages and scores highly on coding and reasoning benchmarks.

• The largest Qwen3 model (Qwen-3-235B-A22B) edges out OpenAI’s o3-mini and Google’s Gemini 2.5 Pro. However, it’s not publicly available yet.

Open source, but make it elite

The biggest model available to the public, Qwen3-32B, still holds its own, even beating OpenAI’s o1 model in some coding tests.

Qwen3 can also be accessed through cloud providers like Fireworks AI and Hyperbolic.

Experts point out that open models like Qwen3 show that open-source AI is keeping pace with closed systems, even as chip restrictions between the US and China tighten.

When you train on 36 trillion tokens ... you flex a little, and that’s okay.