Alibaba has announced a new breakthrough in the field of artificial intelligence. Presented on April 22, 2026, the Qwen 3.6-27B model, despite having 27 billion parameters, outperformed the much larger 397-billion-parameter Qwen 3.5 model in coding and agentic tasks. This innovation expands the possibilities of using AI on standard user computers without the need for massive servers. Ai-stat.ru reports on the development.
The key feature of the new model is its return to a dense architecture. Unlike the Mixture of Experts (MoE) approach that has become popular in recent years, Qwen 3.6 utilizes all 27 billion parameters for every request. This makes the system easier to deploy and its performance more stable. Additionally, thanks to Thinking Preservation technology, the model retains its reasoning throughout a conversation, reducing errors when working with long blocks of code.
The Qwen 3.6-27B model can process text, images, and video through a single encoder. Its context window size is 262,000 tokens, which is sufficient for analyzing an entire software project. In Terminal-Bench 2.0 tests, this model showed results nearly identical to giants like Claude 4.5 Opus. Experts note that developers can now run powerful code assistants on their own personal graphics cards without subscribing to cloud services.
Read “Zamin” on Telegram!
