Qwen3.5-397B-A17B is the first open-weights release in Alibaba's Qwen3.5 family, positioned by its creators as a native vision-language foundation model aimed at reasoning, coding, agent-based tasks, and multimodal understanding including image, video, and GUI interaction. The model is built on a hybrid architecture that fuses linear attention through Gated Delta Networks with a sparse mixture-of-experts design, allowing it to carry 397 billion total parameters while activating only 17 billion per forward pass. Language coverage was expanded from 119 to 201 languages and dialects, broadening the model's practical reach for global developers and enterprises. Weights are published on Hugging Face under Apache License 2.0 and are compatible with Transformers, vLLM, SGLang, and KTransformers, making self-hosting straightforward for teams with suitable hardware.
Beyond the base model, Qwen3.5-Plus serves as a hosted counterpart on Alibaba Cloud Model Studio, offering a one-million-token default context, built-in tools, and adaptive tool use for agent workflows. The open checkpoint itself emphasizes inference efficiency, pairing the sparse expert routing with linear attention so the large parameter budget does not impose proportional compute or latency costs during a single forward pass. In practice, the model is a fit for teams that want to run a strong multimodal foundation locally while retaining the option to scale to the managed Qwen3.5-Plus tier for longer context, integrated tooling, and agent orchestration, all anchored by the same underlying Qwen3.5-397B-A17B design.
