Alibaba releases Qwen3.5, its first natively multimodal open-weight model

Alibaba launched Qwen3.5-397B-A17B, a mixture-of-experts model that handles text, images, and up to two hours of video in a single architecture. Only 17B of its 397B parameters activate per query, making it 19x faster than Qwen3-Max and 3.5-7x faster than Qwen3-235B. The open-weight model tops benchmarks in instruction following and multilingual tasks but still trails GPT-5.2 and Claude 4.5 Opus in reasoning and coding. Alibaba released it under Apache 2.0 as China's open-weight AI race intensifies ahead of DeepSeek's next platform release.

View full digest for February 17, 2026