Plus, 35B-A3B and a preview
Alibaba’s Qwen team opened the Qwen3.6 cycle in two close steps. On 2 April 2026 Qwen3.6-Plus debuted, hosted on Alibaba Cloud Model Studio, with a stated focus on autonomous agents, repository-level coding and multimodal reasoning. On 17 April Qwen3.6-35B-A3B arrived as the first open-weights variant of the family, published on Hugging Face and ModelScope under an Apache 2.0 licence. In parallel, Qwen3.6-Max-Preview is available as a hosted early look at a larger variant on agentic coding, world knowledge and instruction following.
Stated capabilities
Qwen3.6-Plus keeps a 1-million-token context window as the default. Alibaba’s messaging emphasises two directions: software agents capable of working at repository scale — not just local completion — and more accurate multimodal perception on image and video reasoning tasks. Both thinking and non-thinking modes are supported, in line with the previous generation.
Open and accessible
Having an open-weights variant available seventeen days after the Plus launch is a deliberate signal: the family is meant to be shipped both as a managed service and as a downloadable model. The 35B-A3B lives on Hugging Face Hub and ModelScope, while Plus remains the commercial front-end for teams that prefer not to operate the infrastructure. Max-Preview is still rolling out gradually.
What this means in practice
For teams planning on-premise AI infrastructure, the relevant signal is cadence: Alibaba is shortening the gap between hosted launch and open release, and this time did so already mid-month rather than batching at quarter-end. The picture is not yet closed — more sizes of the family are expected — but the “hosted first, open right after” pattern is by now consistent.
Link: Qwen3.6-35B-A3B on Hugging Face · QwenLM/Qwen3.6 repository on GitHub