Krux
February 17, 2026
Alibaba Drops 397B Open-Weight AI Model That's Actually Tiny
Published: February 17, 2026 at 12:25 AM
Updated: February 17, 2026 at 12:25 AM
100-word summary
Alibaba just released Qwen3.5-397B-A17B, the first open-weight flagship in its Qwen3.5 family. Despite boasting 397 billion parameters, only 17 billion activate per task—making it wildly efficient. The model uses hybrid sparse Mixture-of-Experts with linear attention for blazing inference speeds and natively handles text, image, and video inputs. It supports 201 languages, offers a million-token context window via Qwen3.5-Plus, and runs 60% cheaper with 8x efficiency versus predecessors. Weights are live on HuggingFace and ModelScope. This could democratize advanced multimodal AI for developers, putting enterprise-grade agentic capabilities into open-source hands at fraction of the cost.
What happened
Alibaba just released Qwen3.5-397B-A17B, the first open-weight flagship in its Qwen3.5 family. Despite boasting 397 billion parameters, only 17 billion activate per task—making it wildly efficient. The model uses hybrid sparse Mixture-of-Experts with linear attention for blazing inference speeds and natively handles text, image, and video inputs. It supports 201 languages, offers a million-token context window via Qwen3.5-Plus, and runs 60% cheaper with 8x efficiency versus predecessors. Weights are live on HuggingFace and ModelScope.
Why it matters
This could democratize advanced multimodal AI for developers, putting enterprise-grade agentic capabilities into open-source hands at fraction of the cost.