Qwen3.6-35B-A3B: Agentic Coding Power, Now Open to All

Alibaba's Qwen research team has launched Qwen3.6-35B-A3B, a significant open-source challenger in the AI coding assistant space. This 35-billion-parameter model is specifically architected for "agentic" behavior, meaning it can autonomously plan, write, test, and debug code across multiple steps without constant human intervention. Its standout performance is a 91.5% score on the rigorous SWE-bench, a benchmark based on real-world GitHub issues, where it notably outperforms OpenAI's GPT-4 Turbo. This release democratizes high-level coding assistance, offering a powerful, free alternative to closed-source, subscription-based tools.

The model is built on the Qwen3.6 base and fine-tuned with advanced techniques like Direct Preference Optimization (DPO) to excel at following complex technical instructions. It supports a massive 128K context window, allowing it to process extensive codebases and documentation. Available immediately on platforms like Hugging Face and Alibaba's own ModelScope, Qwen3.6-35B-A3B is licensed under the permissive Apache 2.0, enabling both commercial and research use without restrictions. This move pressures proprietary vendors by proving that open-source models can achieve and even exceed the coding capabilities of the best-paid alternatives.