# Qwen

Qwen (Tongyi Qianwen) is Alibaba Cloud's family of predominantly open-weights large and small language models, first launched in beta in April 2023 and known for permissive licenses and wide community use.

Qwen (also known as Tongyi Qianwen, Chinese: 通义千问; pinyin: Tōngyì Qiānwèn) is a family of predominantly open-weights large and small language models (LLM and SLM) developed by [Alibaba Cloud](https://www.wikiprompt.org/wiki/alibaba-cloud). The series includes dense and mixture-of-experts architectures, ranging from small models suitable for edge devices to large-scale models with trillions of parameters. Qwen models are widely used as a base for fine-tuning and abliteration in the open-source community, and a local version powers [Apple Intelligence](https://www.wikiprompt.org/wiki/apple) in China.

The first iteration, based on Meta AI's Llama 1 architecture, launched as a beta in April 2023 and released its 72B weights in December 2023. Subsequent versions, Qwen2 and Qwen3, expanded the family with visual, audio, and omni-modal capabilities, and introduced reasoning-focused models. The latest release, Qwen3.8, includes the 2.4-trillion-parameter Qwen3.8-Max, which was the second largest and second most powerful open-weights LLM and Chinese LLM as of 12 August 2026, after Kimi K3.

## Model Architecture and Training

Qwen models are built on the [Transformer](https://www.wikiprompt.org/wiki/transformer) architecture, initially adopting the [Llama](https://www.wikiprompt.org/wiki/large-language-model) design from Meta AI. Later versions incorporate innovations such as [multi-head attention](https://www.wikiprompt.org/wiki/multi-head-attention), [sparse mixture-of-experts](https://www.wikiprompt.org/wiki/mixture-of-experts) (MoE) layers, and long context windows. For example, Qwen3.8-Max uses a sparse MoE architecture with approximately 95 billion active parameters per forward pass and supports a context window of up to one million tokens. Training data for Qwen3 included 36 trillion tokens across 119 languages and dialects, enabling broad multilingual capabilities.

The family includes both dense models (e.g., Qwen3 dense variants from 0.6B to 32B parameters) and MoE models (e.g., Qwen3-30B-A3B and Qwen3-235B-A22B). Qwen3.8-Max, with 2.4 trillion total parameters, is the largest open-weights model in the series. The models are trained using standard [deep learning](https://www.wikiprompt.org/wiki/deep-learning) techniques, including [Adam](https://www.wikiprompt.org/wiki/adam-optimizer) optimization and [learning rate scheduling](https://www.wikiprompt.org/wiki/learning-rate-schedule), though specific training details are often not fully disclosed.

## Model Variants and Capabilities

Qwen encompasses multiple specialized model lines. The Qwen-VL series combines a vision transformer with an LLM for visual language tasks, with variants such as Qwen2-VL (2B and 7B) and Qwen2.5-VL (3B, 7B, 32B, 72B). Qwen2.5-Omni-7B accepts text, images, videos, and audio as input and generates both text and audio output, enabling real-time voice interaction. Qwen-Audio handles audio tasks, and Qwen2-Math focuses on mathematical reasoning. The QwQ series, starting with QwQ-32B-Preview in November 2024, emphasizes reasoning similar to [OpenAI](https://www.wikiprompt.org/wiki/openai)'s o1, with a 32K token context length.

Qwen3 introduced a thinking mode that can be toggled on or off, allowing the model to deliberate before answering or respond directly. Qwen3.8-Max, released as open weights on 12 August 2026, omits certain cloud features such as image input and non-thinking mode, but the distilled Qwen3.8-27B includes these features and is licensed under the permissive Apache License.

## Releases and Licensing

Alibaba has released over 100 open-weights models, with cumulative downloads exceeding 40 million as of early 2025. The licensing varies by model: many smaller and distilled versions use the Apache 2.0 license, while larger models often require a revenue-sharing agreement for providers generating more than US$50 million annually. For example, Qwen3.8-2.4T-A95B requires such a commercial license, whereas Qwen3.8-27B is Apache 2.0. This strategy balances openness with commercial interests, making Qwen a common starting point for community fine-tuning.

Key release dates include: Qwen beta in April 2023; Qwen 7B weights in August 2023; Qwen 72B and 1.8B in December 2023; Qwen2 in June 2024; Qwen2.5-VL in January 2025; Qwen2.5-Max on 29 January 2025; Qwen2.5-VL-32B-Instruct on 24 March 2025; Qwen2.5-Omni-7B on 26 March 2025; Qwen3 on 28 April 2025; Qwen3.5 in February 2026; Qwen3.8-Max preview in July 2026; cloud release on 3 August 2026; open weights on 12 August 2026; and Qwen3.8-27B on 14 August 2026.

## Ecosystem and Applications

Qwen models are integrated into Alibaba's cloud services and consumer applications. The Qwen mobile app, updated in January 2026, connects to Alibaba's ecosystem, starting with food-service delivery, with plans to extend to platforms like Taobao and Fliggy. A local version of Qwen is used by Apple Intelligence in China, integrated with the operating system. The Accio application, launched in November 2024, leverages Qwen for AI-driven shopping assistance.

In the open-source community, Qwen's permissive licenses have led to numerous fine-tuned and abliterated derivatives, such as "Liberated Qwen" by Abacus AI, which removes content restrictions, and "Qwable", incorporating tuning data from Anthropic's Fable 5. These derivatives highlight Qwen's role as a foundational model for experimentation.

## Recent Developments

In March 2026, Lin Junyang, former head of the Qwen AI model division, resigned after the release of Qwen3.5 and Qwen3.5-Plus, becoming the third executive to leave Alibaba that year. Alibaba reaffirmed its commitment to open-source AI despite these departures. Later that month, a company announcement revealed the formation of a new entity, though details remain limited as of early 2026. The Qwen3.8 series, including the proprietary Qwen3.8-Max and open Qwen3.8-27B, represents the latest advancement, with Alibaba continuing to push the boundaries of open-weights model scale and capability.

---
Source: https://www.wikiprompt.org/wiki/qwen-72b
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-13T03:50:17.531202+00:00