Qwen3.6. This is it.

A developer's viral demonstration shows Alibaba's Qwen3.6-35B-A3B model autonomously coding, testing, and debugging a complete tower defense game. The user gave a single instruction, and the AI agent proceeded to write the code, utilize the Model Context Protocol (MCP) to take screenshots of the installed application for verification, and independently test features like upgrades. Crucially, the model demonstrated meta-cognitive reasoning by identifying its own bugs—noting when the game canvas failed to render and spotting an error in wave completion logic—and then implementing fixes. This was run using the llama.cpp server with a massive 120,384 token context window and specific quantization (Q6_K_XL), highlighting the technical infrastructure enabling such complex, long-horizon tasks.

The event signals a significant shift from AI as a coding assistant to a proactive, executing agent. The model's ability to follow a high-level goal, break it into sub-tasks (coding, screenshot verification, testing, debugging), and course-correct based on observed outcomes mirrors a software engineer's workflow. The developer's shock, noting "I can't imagine what the Qwen Coder that's following will be able to do," underscores the perceived leap in capability. This demonstration was performed with a 35-billion-parameter model, suggesting that highly capable agentic AI is becoming more accessible and efficient, moving beyond pure chat interfaces into actionable, tool-using systems.