- Alibaba open-sourced the 35-billion-parameter Qwen3.6-35B-A3B model, which activates only 3 billion parameters during each inference.
- Excelling in agentic coding and spatial intelligence, the model will help Alibaba gain an edge in AI software and its upcoming smart hardware ecosystem.
Alibaba has open-sourced its large language model Qwen3.6-35B-A3B. As the first open-weight release in the Qwen3.6 series, it continues the company's technical approach of being lightweight and highly efficient.
The highly efficient artificial intelligence model adopts a sparse Mixture-of-Experts (MoE) architecture. While maintaining a lightweight footprint, it demonstrates strong capabilities in agentic coding and multimodal reasoning, according to a Thursday announcement.
The model features 35 billion total parameters but requires activating only 3 billion parameters per inference, significantly lowering computing costs and improving operational efficiency.
Despite its small active parameter size, the model delivers robust performance. It surpasses its predecessor, Qwen3.5-35B-A3B, in agentic coding and rivals larger dense models such as Qwen3.5-27B and Gemma-31B, the announcement said.
It matches mainstream closed-source models on vision-language benchmarks and holds an industry-leading position in spatial relationship understanding.
The model scored 92.0 on RefCOCO and reached 50.8 on ODInW13, according to the statement.
API services will soon be available on Alibaba Cloud's Bailian platform under the name "qwen3.6-flash." It will support the "preserve_thinking" feature, which is tailored to meet the requirements of agentic tasks.
Furthermore, users can seamlessly integrate the model into popular coding assistants like OpenClaw and Claude Code, enabling an efficient and context-aware coding experience, the announcement noted.
Alibaba's open-source small-sized Qwen3.5 models, released last month, previously earned praise from Elon Musk for their "intelligence density."
