Alibaba releases Qwen3.5, an open-weight AI model
Alibaba unveiled Qwen3.5, a 397-billion parameter, open-weight, multimodal Mixture of Experts (MoE) model positioned to compete with OpenAI's GPT-5.2 and Google's Gemini 3 Pro. The model features a hybrid architecture combining linear attention and sparse MoE, activating only 17 billion parameters during inference, which reportedly enables up to 19x faster decoding than previous versions. Running the model requires a minimum of eight GPUs with tensor parallelism for inference and supports a context length of up to 262,000 tokens.
- The "open-weight" release is central to Alibaba's strategy of monetizing AI through its cloud computing division; by making the models accessible, the company drives demand for the cloud infrastructure needed to run and fine-tune them. - This release follows a rapid series of updates, including the Qwen2 family in June 2024, the Qwen3 family in April 2025, and various specialized models for vision (Qwen-VL) and coding (Qwen-Coder). - In benchmark tests for instruction-following accuracy (IFBench), Qwen3.5 scored 76.5, reportedly outperforming systems like GPT-5.2 and Gemini 3 Pro. - The model significantly expands its global reach by increasing its language and dialect support to 201, a substantial increase from the 119 supported in the previous Qwen3 models. - Alibaba is positioning Qwen3.5 for the "agentic AI era," introducing "visual agentic capabilities" designed to allow the system to independently execute tasks across desktop and mobile applications. - The release is part of an intensifying domestic competition among Chinese technology firms, including ByteDance, which recently upgraded its Doubao chatbot, and the influential startup DeepSeek. - A hosted version available through Alibaba Cloud, called Qwen3.5-Plus, offers production-level features including a default context window of 1 million tokens. - Alibaba's open-source Qwen models have become the most widely adopted globally, surpassing 700 million downloads on the developer platform Hugging Face as of January 2026.
