Key Takeaways
- Alibaba released its new Qwen3 family of AI models, stating they compete with top models from Google and OpenAI.
- Many Qwen3 models are available as open-source downloads.
- The models vary in size and use features like “reasoning” for complex tasks and efficient “mixture of experts” architecture.
- Performance tests show Qwen3 models are strong competitors, especially in areas like coding and math benchmarks.
- Google introduced three AI experiments on Google Labs to help people learn languages in personalized ways.
- Google’s experiments use AI for specific learning scenarios like instant phrase help, learning slang, and identifying objects with a camera.
Chinese technology giant Alibaba has launched Qwen3, a new lineup of artificial intelligence models. The company suggests these models can hold their own against, and sometimes even surpass, leading AI from Google and OpenAI.
Most of these Qwen3 models can be downloaded freely under an open license from platforms like Hugging Face and GitHub. They come in various sizes, measured by parameters, which roughly indicate their problem-solving power. Generally, more parameters mean better performance.
The development of models like Qwen by Chinese firms adds to the competitive landscape, pushing American AI labs to innovate further. It has also prompted some regulators to control the export of advanced chips needed for training powerful AI.
According to TechCrunch, Alibaba describes Qwen3 models as “hybrid.” This means they can take extra time to think through difficult problems or provide quick answers for simpler requests. This “reasoning” capability helps them self-check facts, similar to some OpenAI models, but it can make them slightly slower.
Alibaba highlighted that users can adjust how much “thinking” time the models use, tailoring them for specific tasks. Some models also use a “mixture of experts” (MoE) design, which efficiently breaks down tasks for smaller, specialized AI components to handle.
The Qwen3 models support 119 languages and were trained on a massive dataset including textbooks, code, and AI-generated content. Alibaba states these updates significantly boost Qwen3’s abilities compared to its predecessor.
While not necessarily outperforming the absolute best models like OpenAI’s latest, Qwen3 models show strong results. The largest Qwen3 model reportedly beats competitors like OpenAI’s o3-mini and Google’s Gemini 2.5 Pro on coding and math challenges. However, this largest model isn’t yet publicly available.
Even the publicly released Qwen3 models, like Qwen3-32B, are competitive, outperforming older models like OpenAI’s o1 on certain tests. Alibaba also notes Qwen3 is adept at using digital tools and accurately following instructions.
Separately, Google has unveiled three new AI experiments designed to make language learning more personal and engaging. These early-stage tools, potentially challenging established apps like Duolingo, leverage Google’s Gemini AI model.
The first experiment, “Tiny Lesson,” gives you context-specific vocabulary and grammar help for situations like losing a passport. It can even suggest appropriate responses.
Another experiment, “Slang Hang,” aims to help learners sound more natural by teaching colloquial speech and local slang through realistic conversations. Google advises cross-referencing the slang learned here, as the AI can sometimes misuse terms.
The third tool, “Word Cam,” uses your phone’s camera. Snap a picture, and Gemini identifies objects, labelling them in the language you’re learning and suggesting related descriptive words.
Google positions these experiments as explorations into how AI can create more dynamic and personalized self-study experiences. They support numerous languages, including English, Spanish, French, Chinese, and Japanese, and can be found on Google Labs.
