Hacker News discussion covers running the 125B Qwen 3.8 Flash Next model on consumer GPUs (RTX 3090, 4090, 5090) using 3-bit to 5-bit quantization, achieving 40 to 125 tokens per second.

Benchmarks show IQ3_XXS quantization scoring 89.1% on code tasks versus 71.9% for a Q4/Q5 mix, with users reporting fewer hallucinations than 27B 4-bit variants. Initial setup takes about one hour, cold model loading into GPU memory takes roughly 15 minutes, and users compare $1/hour GPU rental costs against $20/month API subscriptions.