Alibaba Launches Qwen3.5 with Visual Agentic AI to Take on US Rivals

Alibaba Enters the Agentic AI Era with Qwen3.5

Alibaba Cloud has released Qwen3.5, its latest flagship AI model, designed for what the company calls the agentic AI era. The model features three hundred and ninety seven billion total parameters but activates only seventeen billion per forward pass through a sparse mixture-of-experts architecture, making it sixty percent cheaper to operate and up to eight times more efficient at processing large workloads than its predecessor.

Qwen3.5 introduces visual agentic capabilities, enabling it to autonomously operate across mobile and desktop applications by analysing screenshots, identifying UI elements, and executing multi-step tasks without human intervention. The model now supports two hundred and one languages, a significant jump from eighty two in the previous generation.

Bold Performance Claims

Alibaba has made striking benchmark claims, comparing Qwen3.5 against OpenAI's GPT-5.2, Anthropic's Claude 4.5 Opus, and Google's Gemini 3 Pro across more than thirty benchmarks. On IFBench, which measures instruction-following ability, Alibaba claims Qwen3.5 outperformed all three. On coding benchmarks it scored eighty three point six on LiveCodeBench and ninety one point three on AIME twenty six for mathematical reasoning. Independent verification of these results remains ongoing.

Ant Group Joins the Push

Alongside Alibaba's release, Ant Group unveiled two trillion-parameter models. Ling-2.5-1T is designed for efficient reasoning with context lengths up to one million tokens, while Ring-2.5-1T is described as the world's first hybrid linear-architecture thinking model. Both achieved impressive results on mathematical reasoning benchmarks, with Ring-2.5-1T reaching gold-tier scores on the International Mathematical Olympiad twenty twenty five.

China's Open-Source Strategy Gains Ground

The releases reflect a broader shift in China's AI industry toward open-source development. Alibaba's Qwen family has overtaken Meta's Llama models in total downloads on Hugging Face, with Qwen derivatives accounting for over forty percent of new language model derivatives on the platform. Both Alibaba and Ant Group have made their latest models available under open licences.