Source · release
Qwen3.5-397B-A17B model card
Summary
The official model card for Qwen3.5-397B-A17B, the first open-weight model in Alibaba's Qwen3.5 series (February 2026). It is a natively multimodal (vision and language) mixture-of-experts model that replaces most attention layers with linear-attention Gated DeltaNet layers. Released under Apache 2.0.
Key points
- 397B total parameters, 17B active; 512 experts, 11 active per token (10 routed, 1 shared).
- Hybrid layout: 60 layers in 15 groups, each with three Gated DeltaNet + MoE blocks and one gated full-attention + MoE block.
- Context: 262,144 tokens native, up to about 1M with YaRN scaling. The hosted Qwen3.5-Plus offers 1M by default.
- Supports 201 languages and dialects (up from 119 in Qwen3).
- Thinks by default before answering; runs on Transformers, vLLM, SGLang, and KTransformers.
- Licence: Apache 2.0.
Why it matters
A leading open-weight model that is mostly linear attention, showing that hybrid architectures are already in production at frontier scale.