February 2026 — Alibaba Cloud has launched Qwen 3.5, its next-generation artificial intelligence model series, positioning it as a major contender in the global race toward autonomous, agentic AI systems. Announced on Monday ahead of the Chinese Lunar New Year, the release introduces open-weight models designed for real-world autonomy, multimodal processing, and significantly lower operational costs, with Alibaba claiming superiority over leading U.S. rivals across key benchmarks.
The flagship open-weight variant, Qwen 3.5-397B-A17B, features 397 billion total parameters but activates only 17 billion per forward pass through an innovative hybrid architecture that combines gated delta networks for linear attention with sparse Mixture-of-Experts (MoE). This design delivers exceptional inference efficiency—decoding speeds 8.6 to 19.0 times faster than its predecessor, Qwen 3-Max—while maintaining high capability. The model includes a native multimodal framework that unifies vision and language processing, trained on trillions of multimodal tokens, supporting text, images, structured inputs, and even videos up to two hours in length.
A standout upgrade is expanded language support, now covering 201 languages and dialects (up from 119 in prior versions), significantly broadening global accessibility. The open variant offers a context window of up to 256,000 tokens, enabling longer reasoning chains and complex document processing.
Alibaba emphasizes Qwen 3.5’s agentic focus: built specifically for “agents of the real world,” it enables the independent execution of complex, multi-step tasks. This includes breaking down goals into subtasks, determining actions, using digital tools across mobile and desktop applications, coordinating workflows, validating outcomes, and proceeding autonomously with minimal human oversight. “Visual agentic capabilities” allow the model to interpret and interact with app interfaces, paving the way for practical automation in enterprise environments, administrative processes, repetitive tasks, and support workflows.
Performance highlights include top rankings on demanding benchmarks. On GPQA (Diamond), it scores 88.7 in graduate-level reasoning (third among evaluated LLMs). On IFBench, it achieves 76.5 in instruction-following precision, outperforming all competitors. Alibaba states the series consistently matches or exceeds models such as OpenAI’s GPT-5.2, Anthropic’s Claude Opus 4.5, and Google’s Gemini 3 Pro, particularly in multilingual, agentic, and multimodal workloads.
Cost efficiency is a core selling point: Qwen 3.5 is 60% cheaper to operate than its immediate predecessor and eight times more efficient at handling large workloads using the same compute resources. Alibaba describes this as setting “a new benchmark for capability per unit of inference cost,” aimed at accelerating adoption among developers and enterprises.
Availability first rolled out through Alibaba’s flagship consumer AI app (Qwen Chat), with official details shared on X by the @Alibaba_Qwen account. The open-weight Qwen 3.5-397B-A17B is accessible on platforms such as Hugging Face, ModelScope, and GitHub (repository: QwenLM/Qwen3.5), while hosted versions like Qwen 3.5-Plus (with up to a 1 million token context window) are offered through Alibaba Cloud Model Studio. Larger Max-series models remain closed for commercial ecosystem use.
The launch intensifies competition in China’s booming AI sector, where Alibaba competes with ByteDance’s Doubao and DeepSeek for chatbot and enterprise leadership. It comes as Chinese firms accelerate open-source and agentic advancements, challenging U.S. leaders in accessibility and cost.
Discussion about this post