What Happened
Alibaba released the first Qwen3.5 model on February 16, 2026. The official repository dates this initial release separately from additional model sizes published on February 24 and March 2. The first downloadable checkpoint was Qwen3.5-397B-A17B, a native vision-language model released under Apache 2.0.
Why It Matters
Qwen3.5 brought visual understanding and general language capabilities into a shared model foundation. Its hybrid design also offered a substantial alternative to relying entirely on conventional attention as model builders sought to reduce the cost of long-context and agent workloads.
Technical Details
The initial mixture-of-experts model contains 397 billion total parameters and activates 17 billion per token. It combines Gated DeltaNet components with gated attention and sparse experts. The developers describe multimodal pretraining and agent-oriented reinforcement learning as core parts of the system. Published weights allow deployment and adaptation outside Alibaba’s hosted services.