Alibaba Launches Qwen3.5 AI Model With 60% Lower Costs, 8x Throughput

Alibaba has officially launched Qwen3.5, the latest version of its flagship artificial intelligence model, positioning it as a system built for the emerging era of AI agents.

The model was released on Feb. 16 in both open-weight and hosted versions, allowing developers to run it on their own infrastructure or through Alibaba Cloud. Alibaba says Qwen3.5 delivers major efficiency gains. According to the company, the model is 60% cheaper to run and offers eight times the throughput when handling large workloads compared with its predecessor.

At the core of Qwen3.5 is a mixture-of-experts architecture. The full model contains 397 billion parameters, but only 17 billion are activated per prompt. That selective activation is designed to reduce hardware costs while maintaining high performance.

According to Alibaba’s technical release, Qwen3.5 combines linear attention mechanisms through Gated Delta Networks with sparse expert routing. The company says this hybrid approach improves inference efficiency without sacrificing reasoning strength.

The model supports up to 262,144 tokens by default, with options to expand further. It also processes text, images, and video natively within a single system. Alibaba says it can analyze videos up to two hours long.

Language support has also expanded significantly, from 119 languages and dialects to 201.

Benchmark claims against US rivals

Alibaba published benchmark comparisons showing Qwen3.5 outperforming earlier Qwen models and, in certain tests, rival systems such as OpenAI’s GPT-5.2, Anthropic’s Claude Opus 4.5, and Google DeepMind’s Gemini 3 Pro.

In IFBench, which measures instruction-following accuracy, Qwen3.5 recorded a leading score of 76.5, according to company data. In other reasoning tests, performance varied, sometimes leading one rival while trailing another.

China’s chatbot war heats up

The rollout comes amid fierce domestic competition. ByteDance released Doubao 2.0 over the weekend, positioning it for agent-based tasks as well. Doubao reportedly commands nearly 200 million users in China.

Meanwhile, industry watchers are awaiting the next-generation model from DeepSeek, whose earlier breakthrough shook global AI markets last year. Alibaba has been aggressively investing in AI infrastructure since that disruption. Bloomberg reported that CEO Eddie Wu pledged more than $53 billion toward infrastructure and AI development.

In its technical overview, Alibaba framed Qwen3.5 as a step toward “universal digital agents.” The company said future development will focus on persistent memory, embodied interfaces for real-world interaction, and systems that can operate autonomously over extended periods.

Related read: Check out this primer on crafting better prompts with AI for smarter interactions and stronger results.