Qwen 3.6 35B A3B

• - 🧠 Sparse mixture-of-experts: 35B total, ~3B active per token.
• - 📏 Native 256K-token context window.
• - 🎯 Tuned for agentic coding, STEM reasoning, and tool use.
• - 🔧 Built-in function calling and web-search capabilities.
• - 🔒 Apache-2.0 licensed, self-hostable open weights.
• - ⚡ Ships as an FP8 checkpoint for efficient inference.
• - 🆕 Flagship open-weight release of the Qwen3.6 generation.
• - 🏢 Developed by Alibaba's Qwen team.

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 3.6 35B A3B is a text model from Alibaba's Qwen team and the flagship open-weight release of the Qwen3.6 generation. It uses a sparse mixture-of-experts design with 35 billion total parameters but only about 3 billion active per token, and natively supports a 256K-token context window. The weights ship under Apache-2.0, including an official FP8 checkpoint, and the model is tuned for agentic coding, STEM reasoning, and tool use.

Relative to its direct predecessor Qwen 3.5 35B A3B, the model keeps the same 35B-total / 3B-active mixture-of-experts layout, positioning the 3.6 release as a successor within the same size class rather than a parameter scale-up. It also sits alongside the smaller Qwen 3.6 27B within the same family.

The model targets developers building coding agents and tool-using workflows, with function-calling and web-search capabilities and open weights that can be self-hosted. As a compact active-parameter MoE, it aims to combine the throughput of a small model with the capacity of a larger expert pool, making it suitable for latency-sensitive agentic and STEM applications.