Qwen3.6

A generation split between open and closed

Qwen3.6 is a generation of the Qwen family of large language models that arrived in April 2026. As with several earlier releases in the line, it was not one model but a set of them, divided between proprietary versions served to customers and open-weight versions that anyone can download.

The first to appear was Qwen3.6-Plus, a proprietary model released on 2 April 2026 and aimed specifically at agentic coding — at writing and running code inside an agent loop, rather than only answering coding questions in a chat window^[1]. Qwen3.6-Plus, together with Qwen3.5-Omni, is catalogued among the proprietary releases of April 2026, while the open-weight Qwen3.6 models were published under the Apache 2.0 licence^[4].

Release timeline

The open-weight line began on 15 April 2026, when Qwen3.6-35B-A3B was posted to Hugging Face as the first open-weight Qwen3.6 model, and continued on 21 April with the dense Qwen3.6-27B^[3][2]. Between the two, on 20 April 2026, came Qwen3.6-Max-Preview, presented as having stronger instruction-following capabilities^[5].

How the open model is built

Qwen3.6-35B-A3B is a mixture-of-experts (MoE) model: it holds 35 billion parameters in total but activates only about 3 billion of them for any given token^[2]. That gap is the point of the design. The network can carry a large store of knowledge while costing roughly what a much smaller dense model costs to run.

The expert layer is cut finely: 256 experts, of which 8 routed experts plus 1 shared expert are active at a time^[2]. The stack is 40 layers deep and mixes two attention designs, Gated DeltaNet and Gated Attention^[2]. It accepts text, image and video input, with a native context window of 262,144 tokens that can be extended to about one million^[2]. The model card lists the weights under Apache 2.0^[2].

Measured results

The model card reports SWE-bench Verified 73.4 for Qwen3.6-35B-A3B, against 70.0 for its predecessor Qwen3.5-35B-A3B, together with SWE-bench Pro 49.5 and Terminal-Bench 2.0 51.5^[2]. The first two measure the repair of real software issues drawn from code repositories; Terminal-Bench measures tasks carried out in a command-line environment. On more general benchmarks it reports GPQA 86.0, MMMU 81.7, AIME26 92.7, LiveCodeBench v6 80.4 and MMLU-Pro 85.2^[2]. These cover graduate-level science questions, multi-discipline multimodal understanding, a mathematics competition set, competitive programming problems and broad knowledge questions respectively.

A design choice aimed at agents

One addition is a `preserve_thinking` option, which keeps the reasoning traces from earlier turns instead of discarding them^[2]. An agent typically holds long multi-turn conversations with tools, and carrying earlier reasoning forward cuts the number of tokens spent working the same ground again^[2].

Where it sits in the line

Qwen3.6 did not stay at the top for long. The proprietary Qwen3.7-Max followed in May 2026 and Qwen3.7-Plus in June 2026^[4]. What the generation illustrates is a release pattern that recurred throughout this period: a closed flagship delivered first to API customers, an open-weight counterpart issued weeks later under a permissive licence, and a size class in the tens of billions chosen so that outside groups could actually run the open model.