The announcement comes with a lot of detail about upgrades in pretraining, multilingual support, agentic thinking, and more. Let’s break it down into simple terms.

Qwen, the AI research team backed by Alibaba, has officially rolled out Qwen3, the newest addition to its growing lineup of large language models. It was announced today, 29 April 2025, and it looks like Qwen is getting serious about shaking up the AI race. With big improvements in coding, maths, general reasoning, and multilingual abilities, Qwen3 is aiming to go head-to-head with models like DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro.

A full spread of models, from small to gigantic

Both dense and MoE (Mixture of Experts) models have been released. Two MoE models and six dense models are now open-weighted and available under an Apache 2.0 license. These can be found on Hugging Face, ModelScope, and Kaggle.

The team recommends using frameworks like SGLang and vLLM for deployment, and Ollama, LMStudio, MLX, llama.cpp, and KTransformers for local use.

Qwen3 is not just one model. It is an entire family. It includes lightweight models like Qwen3-0.6B for small devices and goes all the way up to the massive Qwen3-235B-A22B model. The flagship model packs 235 billion total parameters and uses 22 billion at a time. In comparison, even the tiny Qwen3-4B can rival the older Qwen2.5-72B model in many tasks.

Smarter thinking with two modes

One of the coolest new tricks Qwen3 brings is “hybrid thinking modes.” You get a choice between “thinking mode” and “non-thinking mode.” If the question is tough, the model can slow down and reason step-by-step before answering. If the question is simple, it can shoot back a quick reply without wasting time.

The idea is to give users better control over how much thinking and computing power the model spends. The team said this helps in keeping a balance between cost and speed. It feels like having a sports car that knows when to cruise and when to hit full throttle.

Strong multilingual skills

Qwen3 now supports 119 languages and dialects. From Hindi, Bengali, Marathi, Tamil, to Spanish, Russian, Arabic, and Japanese, it covers a massive range. Whether you are chatting in Assamese, Finnish, or Swahili, Qwen3 can understand and respond.

What the Qwen team says about the future

“We believe we are transitioning from an era focused on training models to one centered on training agents,” the Qwen team said in their announcement. They see Qwen3 as a step toward Artificial General Intelligence and Artificial Superintelligence.

Looking ahead, Qwen plans to push even harder, improving model architecture, scaling data, and developing smarter agentic systems that can reason better and adapt faster.

Major upgrades in training

The pretraining for Qwen3 was a massive task. The model was trained on around 36 trillion tokens, almost double the size used for Qwen2.5. The data comes from websites, documents, and even synthetic maths and coding content generated by older Qwen models.

The training was done in three stages, including a final step where the model’s context length was expanded to handle up to 32K tokens. This allows it to process longer conversations, papers, and complex tasks without breaking down.

Read More:

Greater Noida Authority changes property registration rules