In the field of open-source artificial intelligence, domestic models have once again delivered an impressive performance. On the evening of April 19th, the highly anticipated medium-sized model of the Qwen 3.6 series, Qwen3.6-35B-A3B, was officially released to developers worldwide as open source.

This model's most core feature is its "small but powerful" efficiency. Although the total number of parameters reaches 35 billion, thanks to an advanced Mixture of Experts (MoE) architecture, only 3 billion parameters are activated during actual inference. This means developers can achieve significantly better intelligent output capabilities than models of similar size with lower computing costs.

image.png

In various authoritative benchmark tests, the new model has demonstrated "level-crossing challenges". Whether it is Terminal-Bench 2.0, which evaluates terminal programming capabilities, or assessments targeting real-world agent capabilities, its performance not only far surpasses previous generations but can also compete with larger parameter dense models.

Aside from its strong programming and logical abilities, the model also introduces a "multimodal thinking" mode. When handling complex visual language tasks, it can perform spatial intelligent analysis and recognition of images like humans. High scores in complex image recognition tests such as RefCOCO demonstrate its great potential in understanding the real physical world.

To make technology quickly convert into productivity, the model has achieved deep compatibility with mainstream agent frameworks such as OpenClaw and Claude Code. This high degree of adaptability makes it a preferred foundation for developers to deploy an "intelligent brain" locally, easily handling long-term and complex business logic.

Currently, interested developers can directly download and experience this latest open-source achievement through Mogu Community, Hugging Face, or Qwen Studio