Qwen3.6-35B-A3B recreates 3D scenes from screenshots

Qwen3.6-35B-A3B proves that "small" active parameter counts in MoE architectures can deliver top-tier multimodal spatial reasoning.

• –Recreating 3D scenes from 2D images validates its high RefCOCO spatial scores and its ability to reason about depth, perspective, and geometry.
• –The model’s 1M token context window and "Preserve Thinking" feature allow it to handle complex, iterative 3D asset generation and refactoring across multi-turn sessions.
• –Its performance on QwenWebBench's 3D category demonstrates deep, structured knowledge of modern graphics libraries like Three.js and React Three Fiber.
• –Open-weights availability under Apache 2.0 makes this level of generative spatial intelligence accessible for local deployment on consumer-grade hardware.