Chinese AI giants launch open models ahead of Lunar New Year

What happened

Alibaba and Ant Group have announced a new wave of open artificial intelligence models in the days leading up to the Lunar New Year. The Chinese tech giants are aiming to compete directly with US leaders by focusing on efficiency, advanced reasoning and open code.

Alibaba released its Qwen3.5 model on 16 February, while Ant Group introduced Ling-2.5-1T and Ring-2.5-1T on 15 February. The timing coincides with the one-year anniversary of DeepSeek’s breakthrough, which challenged the idea that only massive budgets can deliver frontier AI systems [Unverified].

Qwen3.5: efficiency and agents

Qwen3.5 is a 397‑billion‑parameter model that, thanks to a sparse mixture‑of‑experts architecture, activates only 17 billion parameters per forward pass. According to Alibaba, this design makes the model about 60% cheaper to operate and up to eight times better at handling heavy workloads than its predecessor.

The model now supports 201 languages, a major jump from the previous generation, and adds what Alibaba calls “visual agentic capabilities”, allowing it to autonomously perform tasks across mobile and desktop apps. The company frames Qwen3.5 as a foundation for the agentic AI era, helping developers and enterprises move faster without increasing compute budgets.

In internal tests, Alibaba compared Qwen3.5 to models such as GPT-5.2, Claude 4.5 Opus and Gemini 3 Pro across more than 30 benchmarks. It claims Qwen3.5 outperformed them on IFBench, which measures instruction-following, though independent verification of these results is still ongoing.

Ling-2.5-1T and Ring-2.5-1T: math and reasoning

Ant Group’s Ling-2.5-1T is the latest trillion‑parameter large language model in its Ling series, designed for higher reasoning efficiency and context windows up to one million tokens. On the AIME 2026 benchmark, the model matched frontier thinking systems while using about 5,890 tokens, compared to the 15,000 to 23,000 tokens typically required.

Ring-2.5-1T is described as the world’s first hybrid linear‑architecture thinking model, built for advanced reasoning tasks. It achieved gold‑tier results on academic math benchmarks, including a score of 35 out of 42 on the 2025 International Mathematical Olympiad and 105 out of 126 on the 2025 Chinese Mathematical Olympiad, surpassing China’s national team cutoff.

Both models are released under open licenses on platforms such as Hugging Face and ModelScope, giving organizations and developers a way to test and fine‑tune them for specific domains and applications.

China’s open-source AI strategy

These launches highlight a broader trend toward open‑source development in China’s AI sector. On Hugging Face, Alibaba’s Qwen family has overtaken Meta’s Llama models in total downloads, according to MIT Technology Review.

By April 2025, new models derived from Qwen made up more than 40% of new language derivatives on the platform, while Llama’s share had fallen to about 15%. A Global Times editorial noted that Chinese companies are using open‑source strategies to lower barriers for global developers and businesses to access advanced AI.

The competitive pressure is visible in Silicon Valley as well: Andreessen Horowitz partner Martin Casado estimates that roughly 80% of startups using open‑source stacks now run on Chinese open models. For many teams, these systems have become the default starting point for building new AI products.

Why it matters

• More efficient open models can lower AI costs and make advanced tools accessible to smaller companies and teams.
• Strong math and long-context reasoning open doors for use in finance, engineering, science and education tools.
• The rise of Qwen and Ant Group’s models strengthens China’s role as a central player in the open-source AI ecosystem.
• Users and builders gain more alternatives to proprietary models, with greater choice over licensing, customization and data control.