Yup agreed, I've ran 1 billion tokens through with >3000 steps and multi-hour autonomous output with nothing more than a 5090, Deepseek Harness, and Qwen 3.8 27

Yup agreed, I've ran 1 billion tokens through with >3000 steps and multi-hour autonomous output with nothing more than a 5090, Deepseek Harness, and Qwen 3.8 27b Quasar NVFP4. You get 260k context, ~150 token/s continuous performance, and full agentic capabilities.