Qwen3.6 35B A3B

Qwen3.6-35B-A3B is a sparse mixture-of-experts language model from Alibaba's Qwen team that pairs 35 billion total parameters with only about 3 billion active per token. The architecture is positioned as a successor to the earlier Qwen3.5-35B-A3B release, and is reported to support multimodal thinking and non-thinking modes alongside its core text generation. Released under the Apache 2.0 license, the weights are openly available, which lets the model be downloaded and run on consumer hardware rather than only through hosted services.

The headline strength of this release is agentic coding: an independent review reports a 73.4% score on SWE-Bench Verified, a level that the same source compares favorably against much larger dense models such as a 31B competitor scoring around 52% on the same benchmark. The model is also exposed through Qwen Studio and Alibaba Cloud's API as a Flash-tier endpoint, giving developers a choice between self-hosted local inference and a managed interface. For practitioners, the practical fit is a coding-oriented assistant that can run on a laptop while still delivering benchmark results competitive with substantially bigger dense alternatives.