Alibaba Cloud, the cloud computing arm of Alibaba Group, launched its next-generation open artificial intelligence model, Qwen-3.5, on February 16, 2026. The company claims the model can compete "with state-of-the-art leading models." The launch was announced on social platform X around 10:00 AM GMT, following an initial release on Alibaba's flagship consumer AI app on Monday afternoon.
The new Qwen-3.5 series introduces significant architectural advancements. A key variant, Qwen3.5-397B-A17B, is described as the first open-weight model in the series, featuring native multimodal capabilities and training for real-world agents. It utilizes a hybrid linear attention + sparse Mixture-of-Experts (MoE) architecture and large-scale reinforcement learning environment scaling. The company claims it achieves an 8.6x to 19.0x increase in decoding throughput compared to its predecessor, Qwen3-Max.
Technically, Qwen-3.5 is built on a unified vision-language foundation trained on trillions of multimodal tokens, enabling it to process text, images, and structured inputs within a single framework. The development employed Gated Delta Networks with sparse MoE routing to increase throughput, limit latency, and reduce computing costs. The model now supports 201 languages and dialects. An open-source variant, Qwen-3.5-Open-Source, was also released, featuring 397 billion parameters and a 256,000-token context window.
Despite the technological announcement, Alibaba Group's stock faced downward pressure. Following the announcement, the company's shares fell 2.93% in global trading, with the stock quoted at $25,160.00 as of February 13. Another report noted a 1.89% decline to close at $155.73 on the same date. The decline extended to a five-day drop of nearly 6%. The launch occurred during a week when Chinese equity markets were closed for the Lunar New Year holiday (February 16-23).
Alibaba positioned Qwen-3.5 directly against leading U.S. AI models. Performance charts shared by the Qwen development team compared it to GPT-5.2, Claude Opus 4.5, and Google's Gemini 3 Pro. In the GPQA Diamond benchmark for graduate-level reasoning, Qwen-3.5 scored 88.7, placing third. However, in the IFBench test for instruction-following accuracy, it recorded a score of 76.5, reportedly outperforming all other models.
The launch is part of a broader competitive push in China's AI sector, where firms are emphasizing open-weight models. This week also saw ByteDance debut its Seedance 2.0 model for video generation and Alibaba's DAMO Academy launch RynnBrain, an AI model for robotics. Alibaba aims to expand its Qwen chatbot application in a domestic market currently dominated by ByteDance's Doubao, which serves nearly 200 million users, and challenged by DeepSeek.
The company stated that Qwen 3.5 operates 60% cheaper than its predecessor and processes heavy workloads eight times more efficiently. It integrates visual agentic capabilities across applications. Alibaba continues to integrate AI into its cloud services and retail operations, recently running a coupon campaign inside the Qwen chatbot that increased active users sevenfold.
