# Qwen3 Qwen3 is a family of large language models developed by Alibaba Cloud, released in 2025. It includes dense and mixture-of-experts variants with open weights, appearing on public leaderboards. Qwen3 is a family of [large language models](https://www.wikiprompt.org/wiki/large-language-model) developed by [Alibaba Cloud](https://www.wikiprompt.org/wiki/alibaba-cloud). Released in April 2025, the family includes both dense and mixture-of-experts (MoE) architectures, with parameter counts ranging from 0.6 billion to 235 billion. The models are notable for their open-weight availability and strong performance on public benchmarks, appearing on media and LLM leaderboards. The release included 20 variants in benchmark snapshots, covering different sizes and configurations. The Qwen3 series builds on the earlier Qwen and Qwen2 model families, which were also developed by Alibaba Cloud. The models are designed for a range of tasks, including text generation, reasoning, and coding. They support multiple languages, with a particular focus on English and Chinese, reflecting Alibaba's global and domestic user base. ## Architecture and Variants Qwen3 models use a [transformer](https://www.wikiprompt.org/wiki/transformer) architecture, the standard for modern large language models. The family includes dense models (where all parameters are active for every token) and MoE models (where only a subset of parameters are activated per token, improving efficiency). The largest dense model has 32 billion parameters, while the largest MoE model has 235 billion total parameters with 22 billion activated per token. The 20 variants in benchmark snapshots include models with 0.6B, 1.7B, 4B, 8B, 14B, 32B dense parameters, and MoE models with 30B-A3B, 57B-A14B, 235B-A22B configurations (where the second number indicates activated parameters). Each size is available in base and instruction-tuned versions, with some also having a "thinking" mode that enables extended reasoning before answering. ## Training and Features Qwen3 models were trained on a large corpus of text data, though Alibaba Cloud has not disclosed the exact dataset size. The training process used standard techniques such as [Adam optimizer](https://www.wikiprompt.org/wiki/adam-optimizer) and [learning rate scheduling](https://www.wikiprompt.org/wiki/learning-rate-schedule). The models incorporate features like [multi-head attention](https://www.wikiprompt.org/wiki/multi-head-attention) and [positional encoding](https://www.wikiprompt.org/wiki/positional-encoding), which are common in transformer-based language models. A key feature of Qwen3 is its hybrid thinking mode. Users can toggle between a fast, direct response mode and a thinking mode that generates internal reasoning steps before producing the final answer. This is similar to approaches used by other model families like [OpenAI](https://www.wikiprompt.org/wiki/openai)'s o1 series, though Qwen3's implementation is open-source. The instruction-tuned versions were aligned using techniques such as [RLHF](https://www.wikiprompt.org/wiki/rlaif) (Reinforcement Learning from Human Feedback) and supervised fine-tuning. The models also support [top-p sampling](https://www.wikiprompt.org/wiki/top-p-sampling) and [temperature scaling](https://www.wikiprompt.org/wiki/temperature-scaling) for controlling output randomness. ## Performance and Benchmarks Qwen3 models have appeared on public LLM leaderboards, including the LMArena (formerly Chatbot Arena) and various academic benchmarks. The largest model, Qwen3-235B-A22B, has shown competitive performance against leading closed models from [OpenAI](https://www.wikiprompt.org/wiki/openai), [Anthropic](https://www.wikiprompt.org/wiki/anthropic), and [Google DeepMind](https://www.wikiprompt.org/wiki/google-deepmind) on tasks like mathematics, coding, and general knowledge. In benchmark snapshots, the 20 variants cover a range of sizes, allowing users to choose a model that balances performance and computational cost. The smaller models (0.6B to 8B) are suitable for edge deployment, while the larger models require substantial GPU resources. The models have been tested on standard benchmarks such as MMLU, HumanEval, and GSM8K, though specific scores vary by variant. ## Availability and Ecosystem Qwen3 models are released under an open license, allowing commercial and research use. They are available for download from Hugging Face and other model repositories. Alibaba Cloud also offers the models through its [cloud platform](https://www.wikiprompt.org/wiki/alibaba-cloud), with APIs for deployment. The models are supported by a growing ecosystem of tools and libraries, including vLLM and Ollama for local inference. They can be fine-tuned using frameworks like Hugging Face Transformers and [PyTorch](https://www.wikiprompt.org/wiki/pytorch). The open-weight nature has led to community adaptations, including quantized versions that run on consumer hardware. ## Reception and Impact The release of Qwen3 was notable for its scale and openness. At the time of release, it was one of the largest open-weight model families available, competing with models from [Meta](https://www.wikiprompt.org/wiki/meta)'s Llama series and [Mistral](https://www.wikiprompt.org/wiki/mistral) AI. The inclusion of MoE variants made large-scale models more accessible, as they require fewer computational resources during inference. Media coverage highlighted Qwen3's strong performance on reasoning tasks and its multilingual capabilities. The models have been used in various applications, including chatbots, code assistants, and educational tools. However, as with all large language models, there are concerns about potential biases and misuse, which Alibaba Cloud has addressed through safety fine-tuning and usage guidelines. As of mid-2025, Qwen3 continues to be updated, with Alibaba Cloud releasing patches and additional variants. The model family represents a significant contribution to the open-source AI community, providing a high-performance alternative to proprietary systems. --- Source: https://www.wikiprompt.org/wiki/qwen3 License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/) Last updated: 2026-09-13T18:56:40.007978+00:00