Qwen 3.6 beats Gemma 4 in code tests

Alibaba's Qwen 3.6 35B is narrowing the gap with frontier models by focusing on agentic efficiency and better logic retention in long contexts. Improved instruction following addresses the "dumb point" issues of previous versions, making it more reliable for multi-file refactors. The 35B-A3B Mixture-of-Experts architecture delivers blazing fast inference, optimized for high-end consumer hardware like the RTX 5090. While Gemma 4 27B/31B often wins on general reasoning, Qwen 3.6 shows superior handling of loosely typed JavaScript and repository-level analysis. Native "Thinking Preservation" mode helps maintain context over a 1M token window, a critical feature for large-scale codebase analysis. Despite progress, the model still occasionally hallucinates non-existent API methods, requiring human oversight for final implementation.