On February 17, just before the Lunar New Year, Alibaba released Qwen3.5 — an open-weight model with 397 billion parameters that expands language support from 82 to 201 languages.
Key upgrades
Native multimodal processing: Text, images, and video are handled within a single model rather than through an external adapter. This approach is similar to Llama 4's early fusion strategy, and native fusion delivers noticeably better results in practice than post-hoc concatenation.
201 languages and dialects: This is likely the broadest coverage among open-source models. For teams building international products, a single model can handle multilingual needs across global markets without requiring separate deployments for each language.
Benchmark performance
Alibaba's own benchmark comparisons show the model matching current offerings from OpenAI, Anthropic, and Google DeepMind. However, CNBC specifically noted that these comparisons are "self-reported data, not independently verified" — a disclaimer rarely seen in Chinese model releases, suggesting overseas media are becoming more cautious about benchmark claims from Chinese AI companies.
What's next
- Qwen3.5-Omni: Supports speech generation in 36 languages. This version has not been open-sourced, breaking Alibaba's previous open-source tradition.
- Qwen 3.6-Plus (released April 2): Further strengthens automated programming and AI agent capabilities.
One signal worth watching: CNBC's report noted that Alibaba, ByteDance, and Zhipu AI all released upgrades in quick succession during the same period, and their shared direction is shifting from chatbots to AI agents. This is not a single company's strategic adjustment but a collective pivot across China's entire AI industry.
Agent capabilities are replacing pure conversation quality as the new main battleground in model competition.
Sources: CocoLoop, CNBC report