Qwen 3.5 9B

• 🆕 Compact 9B dense model from Alibaba's Qwen team
• 📏 262K native context, extendable toward roughly 1M tokens
• 🔧 Gated DeltaNet hybrid attention for efficient long-context inference
• 👁️ Multimodal vision support alongside text input
• 🌐 Supports 201 languages with reasoning and function calling
• 🧠 Thinking mode for reasoning-style responses
• 🔒 Apache 2.0 licensed, open weights on Hugging Face
• 🏢 Part of the broader Qwen 3.5 small-model series

Alibaba

Alibaba Group is a Chinese multinational technology company founded in 1999 and headquartered in Hangzhou, Zhejiang. Originally built around e-commerce and cloud computing, Alibaba has become one of the most prolific contributors to open-weight AI research,…

Qwen 3.5 9B is a compact model from Alibaba's Qwen team, released in early 2026 as part of the Qwen 3.5 small-model series. Per the catalog, it offers vision support, covers 201 languages, and supports reasoning ("thinking") mode plus function calling. Its defining feature is a Gated DeltaNet hybrid attention architecture that pairs linear attention with periodic full-attention layers, giving it a 262K native context window extensible toward roughly 1M tokens.

Compared with same-family predecessors, the model targets efficiency over scale. Its Qwen model card describes the hybrid-attention design as a way to make long-context serving cheaper than the quadratic attention used in earlier dense Qwen generations. It contrasts with the older Qwen3 VL 235B vision line, which used the prior architecture and far larger parameter counts.

Within the 3.5 generation it sits below larger siblings such as Qwen 3.5 35B A3B and the flagship Qwen 3.5 397B, while sharing the same Gated DeltaNet foundation. The later Qwen 3.6 27B continues this hybrid-attention lineage.

The weights are openly available under Apache 2.0 on Hugging Face. Independent or vendor-official benchmark figures from a top evaluator were not available, so no performance numbers are reported here.