Alibaba has officially launched its latest AI models, Qwen3. This new family of AI models performs significantly well in language understanding, reasoning, and efficiency, and is almost on par with popular models from OpenAI and Google.
According to Alibaba, Qwen3 delivers strong performance across benchmarks and tasks. It also introduces new architecture choices that enhance both speed and intelligence.
Qwen3 Offers a Range of Powerful Models
The Qwen3 family includes models of various sizes, from lightweight versions with 0.6 billion parameters to massive ones with 235 billion parameters. In AI, parameters represent a model’s internal decision points. Therefore, more parameters generally mean better performance.
Notably, these models are “hybrid” in design. They switch between fast responses and slower, deeper reasoning depending on the task. This is similar to models like OpenAI’s o3, which balance quality and speed.
Alibaba’s Qwen team explained this in a blog post. “We have seamlessly integrated thinking and non-thinking modes,” they wrote. “Users can adjust the reasoning budget based on their needs.”
This means users can control how much processing power the model spends on a task. As a result, Qwen3 offers both flexibility and performance.
Mixture of Experts
Several Qwen3 models also use a Mixture of Experts (MoE) architecture. MoE divides tasks among smaller, specialized expert models.
Also read: Meta Drops Two New Llama 4 AI Models
These “experts” handle different parts of a problem and work together to complete the task.
This architecture reduces the workload on any single part of the system. It also makes the models more efficient, a key advantage when running AI at scale.
This attribute is very essential, given that Sam Altman, CEO of OpenAI, reported that saying please and thank you cost millions.
Read all about it here: Saying ‘please,’ ‘thank you’ to ChatGPT costs millions: Altman
Multilingual and Data-Rich Training
Alibaba trained Qwen3 on a vast dataset containing more than 36 trillion tokens. Tokens are small bits of data used to train AI models.
For comparison, one million tokens equals about 750,000 words.
The dataset included textbooks, question-and-answer pairs, code snippets, AI-generated content, and web data.
This diversity helps the model handle a wide range of topics and styles. Qwen3 also supports 119 languages, making it one of the most multilingual models to date.
Benchmark Scores Tell the Story
Qwen3’s largest model, Qwen-3-235B-A22B, outperforms OpenAI’s o3-mini and even Google’s Gemini 2.5 Pro on several benchmarks. For example:
- On Codeforces (a programming competition platform), Qwen3-235B-A22B leads.
- It also beats o3-mini on AIME, a difficult math benchmark.
- On BFCL, a test for problem reasoning, Qwen3 scores higher as well.
However, this model is not yet available to the public. Instead, Alibaba has released smaller versions like Qwen3-32B.
Even so, Qwen3-32B competes well against other open-source and proprietary models. It even outperforms OpenAI’s o1 in some cases, such as the coding benchmark LiveCodeBench.
Tool Use, Accuracy, and Instruction Following
Alibaba emphasized that Qwen3 is great at calling external tools, following complex instructions, and replicating data formats.
These are key tasks in real-world AI applications, especially for software developers and enterprise users.
In addition to downloadable models, Qwen3 is now available through cloud AI providers like Fireworks AI and Hyperbolic.
This gives developers flexible access to advanced tools without needing to host the models themselves.
