Will China lead the agentic AI race with Qwen3.5?

On 16 February 2026, Lunar New Year eve, Alibaba Group announced the release of Qwen3.5, the newest model in its Qwen (通义千问 tongyi qianwen) family, explicitly designed for what the company calls the “agentic AI era”.

This is more than an incremental update. Qwen3.5 introduces visual agentic capabilities that allow the system not only to converse but also to see what is happening on a phone or computer screen and act autonomously — opening applications, filling out forms, automating spreadsheets and even controlling complex games or simulations.

Alibaba claims the model is 60% cheaper and eight times more efficient at handling large-scale workloads than its predecessor. This is not mere marketing bravado; it reflects a deeper strategic intent: building scalable AI under hardware constraints imposed by US chip export controls.

So what exactly is agentic AI? If the previous era of large language models (LLMs) was defined by chatbots adept at answering questions, the agentic era is about autonomous agents that can plan, use tools, retain context across sessions and act in digital environments without constant human supervision.

The first open-weight model in the series, Qwen3.5-397B-A17B, adopts an innovative hybrid architecture: 397 billion total parameters, but only 17 billion active per forward pass — instead of turning on the entire “brain” every time, the model only activates a small group of specialists per token. This is enabled by a sparse Mixture-of-Experts design combined with Gated Delta Networks for linear attention — a way for the AI to remember the important parts of a long conversation or document without rereading the whole thing every time, so it runs much faster and uses less memory.

The result is a reduction in memory usage of up to 50%, decoding throughput up to 19 times higher for long contexts, and native multimodal support — text, images and video — from the very start of training.

Benchmarks released by the Qwen team suggest strong performance. On Massive Multitask Language Understanding-Pro (MMLU-Pro), Qwen3.5 scores 87.8 — roughly on par with OpenAI’s GPT-5.2 (87.4) and close to Anthropic’s Claude 4.5 Opus (89.5). In vision benchmarks such as MathVision (88.6) and OCRBench (93.1), it reportedly outperforms both GPT-5.2 and Claude.

In agentic evaluations — BFCL-V4, TAU2-Bench and SWE-bench Verified — the model demonstrates competitive coding-agent and Graphical User Interface (GUI)-agent capabilities. The model is not just good at answering questions — it can act like a junior developer or an assistant that actually uses apps and tools to get things done, and it holds up well on realistic tests against other top models.

More importantly, Qwen3.5-Plus, the hosted version on Alibaba Cloud, ships with a default one-million-token context window and support for 201 languages and dialects, up from 119 previously. This is no accident. Alibaba is clearly targeting global markets, including Southeast Asia, where local language coverage and inference costs often determine adoption.

From a geopolitical perspective, the launch is a shrewd response to the reality of sanctions. The US has restricted China’s access to top-tier Graphics Processing Units (GPUs) such as Nvidia’s H100, H200 and B200. Rather than pursuing ever larger, compute-hungry trillion-parameter models like the earlier Qwen3-Max, Alibaba has pivoted toward architectures that “do more with the same compute”.

This mirrors the approach taken by DeepSeek and Moonshot: prioritising algorithmic efficiency and large-scale reinforcement learning for agentic workflows over pure parameter scaling in the style of OpenAI. The outcome is a robust open-weight ecosystem. Today, many of the top models on machine learning hub Hugging Face’s Open LLM Leaderboard are based on Qwen derivatives.

Competition at home is intensifying. ByteDance released Doubao 2.0 just two days earlier, also branding it as “built for the agent era”. DeepSeek, the startup that shocked the industry last year with low-cost open models, is widely expected to unveil a new generation soon.

Alibaba itself has been aggressive: a recent Lunar New Year 3 billion RMB voucher campaign for the Qwen app, which reportedly drove a sevenfold increase in active users, despite temporary outages. Qwen is no longer just a chatbot; it is evolving into a super-app capable of ordering food, booking tickets and paying bills — all mediated by AI agents.

Analytically, Qwen3.5 signals a broader paradigm shift. For years, the prevailing Western narrative held that “China excels at applications while the US dominates frontier research”. That gap is narrowing rapidly — not because China has discovered a magic algorithm, but because of a hyper-competitive ecosystem: hundreds of startups, a vast Science, Technology, Engineering, and Mathematics (STEM) talent pool (producing roughly four times as many STEM graduates annually as the US), and strong state backing.

Alibaba’s open-weight strategy — releasing a 397B-parameter model on Hugging Face and ModelScope — speeds up knowledge sharing and creates powerful network effects that closed models like GPT or Claude struggle to replicate. By letting the global community study, adapt and build on the model, Alibaba fuels a fast-growing ecosystem that proprietary platforms cannot easily match.

Still, it would be premature to declare China the outright winner. Agentic AI remains fraught with pitfalls. The ability to “see screens and act” sounds revolutionary, but in real-world settings, error compounding can be dangerous — imagine an agent misclicking and transferring millions of dollars or automating medical decisions based on biased data. Issues of safety, alignment, and persistent cross-session memory are far from resolved. Benchmarks, too, are often optimised; real-world performance in multi-day tasks, multi-agent collaboration or noisy environments remains largely untested.

Moreover, the agentic era will intensify debates around labour and regulation. In China, where the population is ageing and labour costs are rising, virtual agents could displace millions of administrative and customer service jobs. Similar anxieties exist in the West, compounded by concerns over privacy and Big Tech monopolies. Who will regulate “agent rights”? Should AI be allowed to act on a human’s behalf without explicit consent every time? What happens when Chinese and American agents interact on global platforms under different rule sets?

Alibaba appears aware of these challenges. In an official Qwen blog post, the team wrote: “The next leap requires shifting from model scaling to system integration: building agents with persistent memory for cross-session learning, embodied interfaces for real-world interaction, self-directed improvement mechanisms, and economic awareness to operate within practical constraints.” It is a candid admission that a single model is not enough. What is required is an ecosystem — agent frameworks, memory layers, safety guardrails, and integration with the physical world.

In conclusion, Qwen3.5 is not the end of the story but the opening of a new chapter — one that underscores a simple reality: China is no longer merely chasing from behind. Through efficient architectures, aggressive open-source releases, and a clear focus on agentic capabilities, Alibaba is laying the groundwork for AI that is genuinely affordable and globally scalable. For countries like Indonesia, this represents a golden opportunity to leap forward — provided there is smart regulation, strong local talent and a sovereign data strategy.

The race toward AI supremacy, however, is not a short sprint toward the much-hyped Artificial General Intelligence but a long and demanding systems marathon. In the agentic era — where models like Qwen3.5 can see screens, plan actions and operate autonomously — the central question is no longer how intelligent an agent is, but how safe, trustworthy and aligned it is with societal values. A highly capable agent that is vulnerable to cascading errors — misinterpreting instructions and causing financial losses or dangerous outcomes — can permanently erode public trust.

Poor alignment also risks amplifying systemic bias, deepening social inequities, or triggering cross-cultural value conflicts when Chinese and Western agents interact on global platforms. Alibaba’s aggressive open-weight and cost-efficiency approach has thrown down a serious gauntlet, forcing competitors — from OpenAI to European startups — to focus not only on performance but also on robust guardrails, secure persistent memory and effective human oversight. The challenge is further complicated by the scale at which agentic AI may replace routine work, raising ethical questions about agent rights, legal liability for autonomous actions and data privacy amid dependence on specific cloud infrastructures.

Now that Alibaba has decisively positioned itself as a pioneer of affordable, scalable agentic AI, it is up to the rest of the world — regulators, academics, and companies in developing countries such as Indonesia — to respond with mature strategies. The task is not merely to adopt the technology but to shape it so that it aligns with national interests, human values and long-term sustainability. Only then can the agentic AI era become not a destructive race but a foundation for inclusive and responsible progress.

Related: ByteDance riles smartphone ecosystem with AI pitch | Pudu Robotics founder Zhang Tao: Humanoid robots will be bigger than cars