Qwen
Based on Wikipedia: Qwen
When Alibaba Cloud unveiled Qwen in April 2023, few could have anticipated the impact it would have on the global AI landscape. The company, best known for its e-commerce empire, quietly entered the language model race with a model built atop Meta's Llama architecture—a strategic choice that would prove pivotal.
Alibaba's journey began with Tongyi Qianwen (通义千道), launched as a beta in April 2023 and opened to the public by September of that year after securing regulatory approval. The timing was deliberate: China had just begun rolling out its sweeping generative AI regulations, and Alibaba wanted to be among the first wave of compliant entrants.
The architecture tell a fascinating story. Rather than building from scratch, Qwen borrowed heavily from Meta's Llama—specificallyMeta's Llama design. This wasn't laziness; it was pragmatism. By adapting an already proven foundation, Alibaba could focus resources on scaling and fine-tuning rather than fundamental research. The decision paid off: by December 2023, they released both the 72B and 1.8B parameter models for download, followed by the 7B weights in August.
But here is where things get complicated—and where the narrative gets messy. Alibaba positioned Qwen as "open source," a term that has become almost meaningless in the AI age. The company released model weights under the Apache 2.0 license, allowing anyone to download and use them freely. However, they retained the training code and never documented the training data. This distinction matters enormously.
The Linux Foundation's Open Source AI Definition and Model Openness Framework both distinguish between releasing weights and providing full transparency. Qwen falls short of both definitions—a fact that rarely appears in press releases but is crucial for understanding what "open source" actually means in this context.
By July 2024, benchmarking platform SuperCLUE offered validation. Their ranking placed Qwen2-72B-Instruct behind OpenAI's GPT-4o and Anthropic's Claude 3.5 Sonnet, but ahead of every other Chinese model. The message was clear: Qwen had become China's competitive global AI contender.
The release cadence accelerated through 2024 and 2025—Qwen2 in June, followed by a November surprise that caught the industry off guard. QwQ-32B-Preview, a reasoning-focused model similar to OpenAI's o1, was released under Apache 2.0 with only weights made available (no dataset or training method). It offered a 32K token context length and outperformed o1 on certain benchmarks.
November also saw the launch of Accio, an AI-native application built atop Qwen specifically designed for market insights and sourcing questions within Alibaba's B2B e-commerce ecosystem. The tool automates labor-intensive tasks like data collection and trend tracking—proof that Qwen wasn't merely a research project but a practical business instrument.
The visual language models tell their own story. Qwen-VL combined vision transformers with LLMs, evolving through Qwen2-VL (with 2B and 7B parameters) to Qwen2.5-VL in January 2025, featuring variants of 3, 7, 32, and 72 billion parameters. All except the 72B variant operate under Apache 2.0.
Qwen-VL-Max emerged as Alibaba's flagship vision model, offered through Alibaba Cloud at $0.41 per million input tokens—a pricing structure that reveals their commercial intent alongside open-weight releases.
The numbers tell a story of massive adoption. Over 100 open-weight models released, more than 40 million downloads—a figure that speaks to both community interest and Alibaba's distribution strategy.
Fine-tuned versions by enthusiasts appeared, including "Liberated Qwen" developed by San Francisco-based Abacus AI—specifically designed to respond without content restrictions, a fascinating departure from the company's otherwise careful alignment.
The pace did not let up. January 2025 brought Qwen2.5-Max. March delivered two releases: Qwen2.5-VL-32B-Instruct under Apache 2.0 on the 24th, and Qwen2.5-Omni-7B on the 26th—a multimodal model accepting text, images, videos, audio input and generating both text and audio output, enabling real-time voice chatting.
April 2025 marked the arrival of Qwen3, with all models under Apache 2.0. The family included dense variants (0.6B, 1.7B, 4B, 8B, 14B, 32B parameters) and sparse models (30B with 3B activated, 235B with 22B activated). They were trained on 36 trillion tokens across 119 languages.
September brought Qwen3-Max on the 5th, followed by Qwen3-Next under Apache 2.0 on the 10th—both accessible through chat.qwen.ai and platforms like Hugging Face and ModelScope.
The month ended with Qwen3-Omni on September 22nd—a mixed multimodal model generating text, images, audio, and video under Apache 2.0, available across the same channels.
January 2026 saw Qwen3-Max-Thinking capable of generating text, pictures, or video—released on the 27th.
February brought Qwen-3.5 on the 17th, followed by Qwen3.5 and Qwen3.5-Plus on the 16th—all open-weights.
In early 2026, several Qwen executives resigned, including Lin Junyang who led development of Qwen3-Max and Qwen3.5. The departures sparked concern that Alibaba might shift away from research and open-source AI.
The company responded by reaffirming its focus on open source—an answer to the critics but also a strategic position in an industry increasingly defined by openness.