Krux
February 17, 2026
Alibaba's Qwen3.5 Claims to Beat GPT-5.2 at 60% Less Cost
Published: February 17, 2026 at 11:36 PM
Updated: February 17, 2026 at 11:36 PM
100-word summary
Alibaba just dropped Qwen3.5, an agentic AI model designed to autonomously handle complex tasks across apps, devices, and languages. The beast packs 397 billion parameters but activates only 17 billion per task using a sparse Mixture-of-Experts setup, making it blazing fast and surprisingly cheap. It understands text, images, video, and UI screenshots across 201 languages, with context stretching up to 1 million tokens. Alibaba claims it outperforms GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro while running 60% cheaper and eight times faster on big workloads. This launch signals China's serious play to dominate enterprise AI infrastructure globally, especially as Alibaba pours 380 billion yuan into AI expansion through 2028.
What happened
Alibaba just dropped Qwen3.5, an agentic AI model designed to autonomously handle complex tasks across apps, devices, and languages. The beast packs 397 billion parameters but activates only 17 billion per task using a sparse Mixture-of-Experts setup, making it blazing fast and surprisingly cheap. It understands text, images, video, and UI screenshots across 201 languages, with context stretching up to 1 million tokens. Alibaba claims it outperforms GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro while running 60% cheaper and eight times faster on big workloads.
Why it matters
This launch signals China's serious play to dominate enterprise AI infrastructure globally, especially as Alibaba pours 380 billion yuan into AI expansion through 2028.