Alibaba (Qwen)

What it is

Alibaba Group's large language model family, developed by Alibaba Cloud's Qwen team based in Hangzhou, China. The full Chinese name, Tongyi Qianwen (通义千问), translates loosely as "a unified response to a thousand questions." The family spans dense models from 0.6 billion to 235 billion parameters, vision-language and speech variants, a dedicated coding line (Qwen3-Coder), and MoE (Mixture-of-Experts) architectures scaling to 480 billion total parameters. Qwen functions as the AI intelligence layer of Alibaba Cloud, competing for enterprise and developer customers across China and globally. Most models ship under the Apache 2.0 license; frontier flagships are kept proprietary and sold as an API through Alibaba Cloud.

History

Alibaba launched a closed beta in April 2023. The service opened publicly in September 2023 after receiving clearance under China's Generative AI Service Regulations, which require security assessments for public-facing AI products. The first open-weight release came in December 2023, putting 1.8B and 72B parameter models on Hugging Face. Qwen2 arrived in June 2024; Qwen2.5 followed in September 2024, introducing a MoE architecture that outperformed Meta AI's Llama 3 on several benchmarks. Qwen3, released April 28, 2025, was the most consequential generational step. It introduced hybrid reasoning, a toggle between "thinking mode" for multi-step tasks and standard generation for faster responses, trained on 36 trillion tokens, double the corpus of its predecessor. The Qwen team received the NeurIPS 2025 Best Paper Award for architecture research incorporated into the Qwen3-Next line. Qwen3-235B-A22B scored 95.6 on ArenaHard and 85.7 on AIME'24, competitive with the best closed-source models at that time.

Current state

As of July 2026, Alibaba runs a two-tier model strategy. The open-weight tier centers on the Qwen3.6 series (released April 2026, Apache 2.0); Qwen3.6-27B outperforms the larger Qwen3.5-397B-A17B on SWE-bench Verified (77.2 versus 76.2), a coding-agent benchmark. The proprietary tier, the Qwen3.7 line unveiled at Alibaba's Apsara Summit in Hangzhou on May 20, 2026, consists of Qwen3.7-Max (flagship) and Qwen3.7-Plus (multimodal agent variant, released June 1, 2026), both API-only via Alibaba Cloud. By May 2026 the Qwen app had 234 million registered users and cumulative downloads across the family exceeded 300 million. The open lower tier builds global developer mindshare; the proprietary flagships monetise via cloud subscription, mirroring the playbook used by DeepSeek and, internationally, by OpenAI.

Relationships

Qwen is Alibaba's direct competitive response to DeepSeek, the other dominant Chinese open-weight lab. Among Chinese frontier labs, Qwen is regularly benchmarked alongside MiniMax and Zhipu GLM, though Qwen's parameter scale and download volumes give it greater reach. Internationally, the open-weight line competes most directly with Meta AI's Llama series. The most consequential external event as of mid-2026 is Anthropic's accusation that Alibaba ran the largest known model distillation attack on any AI system: 28.8 million fraudulent exchanges through 25,000 fake accounts between April 22 and June 5, 2026, targeting Claude's software-engineering and agentic-reasoning capabilities for Qwen training. The US Department of Defense has placed Alibaba on its Section 1260H "Chinese military company" list, complicating US enterprise procurement of Qwen-powered services.

What to watch

Whether the dual-tier open/closed strategy sustains developer loyalty as DeepSeek and Meta AI continue releasing fully open frontier models. How Anthropic's distillation complaint, referred to the US Senate in June 2026, shapes US-China AI intellectual-property law and whether Congress legislates against distillation. How US export controls on advanced semiconductors constrain Alibaba Cloud's capacity to train Qwen successors. Whether the open-vs-closed frontier debate shifts Qwen's licensing posture. Progress on the 119-language multilingual roadmap, the clearest product differentiation from English-first Western labs.