Open Source

Alibaba's Qwen3.8-27B crushes BASIC ray-tracer coding

Qwen3.8 autonomously writes a recursive ray-tracer in BASIC without human intervention

Deep Dive

Alibaba's Qwen3.8-27B has demonstrated a significant leap in autonomous coding capabilities, outperforming its predecessor Qwen3.6-27B in generating a functional recursive ray-tracer written in BASIC. The feat was showcased by a Reddit user who built an agentic harness equipped with a BASIC-to-JavaScript transpiler, enabling the AI to write code, execute it, analyze the output, and iterate toward a correct solution.

The prompt tasked the model with creating a ray-tracer that renders three metallic spheres (copper, silver, gold) over a glossy checkerboard plane under a deep blue sky, using the Cook-Torrance lighting model. Qwen3.6 could produce a working ray-tracer but often missed subtle errors that required explicit user feedback to fix. In contrast, Qwen3.8 autonomously iterated to a visually accurate solution without human intervention, showcasing improved self-correction and debugging abilities. Both models were quantized using unsloth's UD-Q8_K_XL, highlighting that the advancements stem from architectural and training improvements rather than hardware or quantization alone.

Key Points
  • Qwen3.8-27B autonomously writes and debugs a recursive BASIC ray-tracer without human input
  • Qwen3.6-27B required manual corrections, while Qwen3.8 iterates to a correct solution independently
  • Both models ran on unsloth UD-Q8_K_XL quants, proving advancements are model-driven

Why It Matters

Qwen3.8's autonomous coding breakthrough signals a major step toward agentic AI that can self-improve without human oversight.

📬 Get the top 10 AI stories daily