- Release of code weights for the giant Qwen model
- Architectural design and superior performance benchmarks
- Differences between the open version and the cloud service
- Compact version and local developer support
- Intense competition among Chinese AI labs
- Frequently asked questions
Release of code weights for the giant Qwen model
Alibaba Cloud has taken an exceptional step in the global artificial intelligence community by releasing the full code weights of its flagship Qwen 3.8 model, featuring 2.4 trillion parameters, for free via Hugging Face. This marks the first time the Chinese company has decided to make a max-tier ultra-capable model open source, allowing developers and researchers worldwide to download and use it in their independent applications.
The model relies on a complex mixture-of-experts architecture, activating only 95 billion parameters per text token during processing and inference, which ensures high computational efficiency and seamless compatibility with the latest globally adopted programming inference frameworks such as vLLM and SGLang.
Architectural design and superior performance benchmarks
Approved benchmark test results showed exceptional superiority for the new Qwen model. On the OSWorld benchmark, designed to evaluate models’ ability to perform independent tasks and control computer systems, the model achieved 86.1 points, outperforming Anthropic’s Claude 5, which scored 85 points.
The independent evaluation platform Artificial Analysis also ranked it at an advanced intelligence level between OpenAI’s flagship GPT-5.6 Sol and GPT-5.6 Terra models. The original model supports a massive context window of 262,000 text tokens, along with multimodal processing for text, images, and video clips, achieving record inference speeds exceeding 4,000 tokens per second per processing unit on advanced NVIDIA hardware.
Differences between the open version and the cloud service
This open release comes just a few days after Alibaba launched the commercial version of the model via its Model Studio cloud platform at $2 per million input tokens and $6 per million output tokens. The commercial cloud version is capable of expanding its context window to completely exceed one million text tokens.
In contrast, technical reports clarified that the open weights published on Hugging Face are designed for text processing and come under a new usage license that includes revenue-sharing clauses for major commercial enterprises that achieve massive commercial uses of the model, balancing the promotion of open innovation with the preservation of the company’s investment returns.
Compact version and local developer support
Alibaba intends to enhance its open-source ecosystem by releasing a smaller version, Qwen 3.8 with 27 billion parameters, on August 15. This version is specifically designed for developers and medium-sized enterprises wishing to run highly efficient artificial intelligence models on local devices and servers without needing a giant cloud computing infrastructure.
This step reflects the overwhelming success achieved by the Qwen model series, whose total downloads on Hugging Face have surpassed the one billion mark, making it the most widespread and adopted ecosystem among developers in Asia and around the world.
Intense competition among Chinese AI labs
Alibaba’s launch comes amid a fierce technological race inside China, where the market is witnessing intense competition with major national laboratories such as Moonshot AI, which launched the Kimi K3 model with 2.8 trillion parameters with partial backing from Alibaba itself, alongside models from ByteDance and Tencent.
This momentum underscores the ability of Chinese companies to deliver artificial intelligence models that rival their American counterparts in efficiency and intelligence, while cementing China’s presence as a superpower in leading the global open-source software movement.
Frequently asked questions
Question: What is the total parameter size of the new Qwen 3.8 model?
Answer: The model’s total size is 2.4 trillion parameters, with 95 billion parameters activated per text token using the mixture-of-experts system.
Question: How did Qwen outperform Anthropic models in computer control tests?
Answer: The model scored 86.1 points on the OSWorld benchmark, surpassing Claude 5, which scored 85 points.
Question: When will the compact version for local operation of the model be launched?
Answer: The Qwen 3.8 model with 27 billion parameters will be launched on August 15 to facilitate running on local devices.
