Qwen

• 1. The First Public Release

  ›

  Alibaba Cloud launched a beta version of Qwen in April 2023 under the name Tongyi Qianwen. The system opened for public use in September 2023 after receiving regulatory clearance from Chinese authorities. This initial release marked the beginning of a rapid development cycle that would see multiple major updates within just two years. The model architecture was based on the Llama architecture developed by Meta AI, establishing a foundation for future iterations. In December 2023, Alibaba released its 72B and 1.8B models for download to the public. These early releases included weights but did not provide training code or documented datasets.

  568

• 2. Architectural Shifts And Variants

  ›

  Qwen2 arrived in June 2024 with both dense and sparse model configurations available to developers. By November 2024, the company introduced QwQ-32B-Preview, a specialized reasoning model similar to OpenAI's o1. This variant featured a 32K token context length and outperformed competing systems on certain benchmarks despite releasing only weights without dataset documentation. The Qwen-VL series combined vision transformers with large language models to process visual data alongside text. In January 2025, Qwen2.5-VL expanded this capability with variants containing 3, 7, 32, and 72 billion parameters. The flagship vision model Qwen-VL-Max sold through Alibaba Cloud at US$0.41 per million input tokens as of 2024.

  24707981

• 3. Open Weights Versus Proprietary Access

  ›

  Many Qwen variants distributed under Apache-2.0 licenses while others remained proprietary cloud offerings. Alibaba released more than 100 open weight models that had been downloaded over 40 million times by early 2025. Fine-tuned versions like Liberated Qwen emerged from enthusiasts such as San Francisco-based Abacus AI. These community adaptations responded to any user request without content restrictions. The training code for most models never reached public release nor did the training data get documented. This approach meant they failed to meet terms of either the Open Source AI Definition or Model Openness Framework from the Linux Foundation. Most advanced capabilities stayed locked behind Alibaba Cloud subscriptions rather than available freely online.

  31354142

• 4. Benchmark Performance And Comparisons

  ›

  South China Morning Post reported in July 2024 that SuperCLUE ranked Qwen2-72B-Instruct behind GPT-4o and Claude 3.5 Sonnet but ahead of other Chinese models. A blog post from the 29th of January 2025 claimed Qwen2.5-Max outperformed foundation models including DeepSeek-V3 and Llama-3.1-405B in key benchmarks. By the 5th of September 2025, Alibaba stated its Qwen3-Max surpassed non-reasoning models like Kimi K2 and Claude 4 Opus Non-thinking. The company announced on the 26th of March 2025 that Qwen2.5-Omni-7B performed comparable to or better than Qwen3-32b while using less than 10% of its training cost measured in GPU hours. Inference throughput reached greater than 10x higher when contexts exceeded 32K tokens for certain configurations.

  738691

• 5. Multimodal Processing And Real-Time Features

  ›

  Qwen2.5-Omni-7B accepted text images videos and audio as input while generating both text and audio output. This capability enabled real-time voice chatting similar to OpenAI's GPT-4o system. The model became available through chat.qwen.ai and platforms like Hugging Face GitHub and ModelScope starting February 2025. Qwen3-Omni released on the 22nd of September 2025 processed mixed multimodal data including natural speech streaming responses. These systems combined vision transformers with language processing to handle complex visual tasks alongside standard text generation. Audio handling capabilities allowed the family to process spoken language directly without requiring separate transcription services. Streaming features delivered responses in both written text and natural human-like speech formats simultaneously.

  294951