Qwen 27B Runs 21 Days on a Single RTX 3090 to Write Its Own CUDA Kernel

A Reddit user ran a Qwen3.8-27B agent on a single RTX 3090 for 21 unsupervised days, during which it built a working CUDA inference engine reaching over 100 tokens/s, with only about 12 human interventions.

2026-09-21 ~ 2026-09-22 · 2 related posts

• One RTX 3090 ran Qwen 27B autonomously for 3 weeks — it shipped working CUDA kernels — skeole · 2026-09-21
• Qwen3.8-27B spent 21 days building its own CUDA engine on one RTX 3090 — DanGrover · 2026-09-22