Alibaba launches Qwen3.6 for agentic coding

Background

Alibaba's Qwen team introduced Qwen3.6 as the next step after Qwen3.5. The release was aimed at agentic coding, a style of programming in which a model plans and carries out multi-step work with tools, and it arrived in two parts over April. ^[1]

What happened

On 2 April 2026, Alibaba released the proprietary Qwen3.6-Plus, aimed at agentic coding. On 15 April, the first open-weight Qwen3.6 model, Qwen3.6-35B-A3B, was posted to Hugging Face, and the dense Qwen3.6-27B followed on 21 April. A Qwen3.6-Max-Preview came on 20 April. ^[1][2][3][5]

Key details

Qwen3.6-35B-A3B has 35 billion total parameters and about 3 billion activated per token. Its architecture mixes Gated DeltaNet and Gated Attention layers with a 256-expert mixture of experts. ^[2]

Among its features is a preserve_thinking option, which keeps reasoning traces from earlier turns to cut token use in agent workflows. ^[2]

Alibaba reported 73.4 on SWE-bench Verified, up from 70.0 for Qwen3.5-35B-A3B, along with 49.5 on SWE-bench Pro, 51.5 on Terminal-Bench 2.0, 86.0 on GPQA and 81.7 on MMMU. These are the developer's own figures. ^[2]

Impact

The release shows the family using both hosted and downloadable channels within one generation. ^[1][2]

Several terms in the release are worth defining. Agentic coding means a model works on programming tasks by planning steps, using tools such as code editors or test runners, and checking its own results, rather than answering a single prompt. SWE-bench asks models to resolve real issues from open-source software projects, and Terminal-Bench measures how well a model completes tasks in a command-line environment. Both are useful for comparing coding ability, but each covers only part of the work that software development involves. Agentic coding is the stated focus of the release, which is why the preserve_thinking option, aimed at multi-turn agent workflows, is relevant.