Qwen3: Alibaba makes new AI model freely available

Qwen3 comes in two versions: a large LLM with the bulky name Qwen3-235B-A22B, which, according to Alibaba, should be able to keep up with models such as DeepSeek R1 o3-mini from OpenAI and Gemini 2.5 Pro in common benchmarks, and a small Qwen3-30B-A3B, which should nevertheless be extremely powerful.

Both are mixture-of-experts models, i.e., they consist of several experts, one or more of which react to an input instead of the entire model being addressed. The numbers and letters in the model names are accordingly the total number of parameters (B) and the individual expert parameters (A). The two large models as well as six small, condensed models are published under Apache 2.0 license.

Alibaba writes on the blog post: “We are convinced that the publication and open sourcing of Qwen3 will significantly advance the research and development of large foundation models. Our goal is to enable researchers, developers, and organizations around the world to develop innovative solutions using these innovative models.”

Qwen3 provides thinking mode or quick answers

Hybrid means that Qwen3 can be used in a “think mode” or without such a process. In thinking mode, it takes longer to answer a query and the model takes time to check the answer. The “non-thinking mode”, on the other hand, produces an answer quickly. However, this may be less accurate or incorrect. AI providers say that it is less “deep”. Without the depth, the answers are definitely more favorable.

Qwen3 supports 119 languages and dialects, including Chinese, English and Spanish, of course, but also German, Luxembourgish and Yiddish. The model is also said to be optimized for agent tasks. The data used to train Qwen3 includes PDF documents as well as content from the web, writes Alibaba. These were combined from the predecessor model Qwen2.5 for training. The math and coding skills were also provided using synthetic data generated by the predecessor model.

There were three levels of training in total. The classic pre-training for basic knowledge, a second training specialized in STEM areas and finally a training on particularly long and high-quality content. According to the benchmark results published by Alibaba, Qwen3 can at least keep up with the largest models currently available from other providers. In practice, however, this means little.

The Qwen-based AI assistant Quark is already the most popular AI service in China. Meta AI is said to be the most used AI chatbot worldwide. However, it is not yet available in China. Apple also does not offer direct access to ChatGPT on its devices in China. Instead, Alibaba's models are to be used there.

(emw)