What happened
Alibaba Cloud released Qwen 3.5 on February 16, 2026 – timing the launch on the eve of Lunar New Year and capping a week where nearly every major Chinese AI developer rolled out new flagship models.
The release includes two versions: an open-weight model with 397 billion parameters available for download, and a hosted version running on Alibaba’s cloud platform Model Studio. According to Alibaba’s announcement, Qwen 3.5 is 60% cheaper to run and offers eight times the throughput compared to its predecessor.
The model uses a mixture-of-experts architecture – while it contains 397 billion total parameters, only 17 billion activate per prompt. This selective activation reduces hardware costs while maintaining performance.
Technical capabilities
Qwen 3.5 was built with what Alibaba calls “native multimodal capabilities” – trained from the ground up to understand text, images, and video simultaneously. The model supports 201 languages and dialects, up from 82 in the previous generation.
The company emphasized agentic features: visual capabilities that allow the model to interact with mobile and desktop applications independently, completing multi-step tasks like research or service bookings without constant human guidance.
Alibaba published benchmark comparisons showing Qwen 3.5 performing on par with GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro in categories like knowledge reasoning and instruction following. These comparisons were self-reported and not independently verified.
Why this matters
The timing is strategic. Alibaba released Qwen 3.5 just days after ByteDance launched Doubao 2.0, which commands nearly 200 million users in China. Both companies positioned their new models as suited for the “AI agent era” – systems that can act independently rather than just respond to prompts.
This release also signals intensifying US-China AI competition. With export controls limiting Chinese access to advanced chips, domestic companies are focusing on efficiency and optimization rather than sheer parameter scale. Qwen 3.5’s architecture reflects this – achieving competitive performance while using only a fraction of total parameters per query.
For developers outside China, the open-weight release matters. Alibaba claims the Qwen family has produced more than 140,000 derivative models worldwide. The expanded language support – particularly for South Asian and African dialects – suggests Alibaba is targeting global adoption, not just domestic market share.
Sam Altman announced on the same weekend that OpenClaw’s creator would join OpenAI, noting that Qwen models are compatible with OpenClaw’s agent framework. The Chinese open-source ecosystem is increasingly interoperable with Western AI tools.
What comes next
Alibaba’s technical lead said the company expects to release more open-weight models during Lunar New Year. The broader race in China now centers on agent capabilities – who can build AI that reliably acts on users’ behalf.
DeepSeek is expected to launch its next-generation model in coming days. After the startup’s viral rise last year triggered a global tech selloff, investors and competitors are watching closely.
