Strata might be actually really detrimental for intelligence? or is it Swift and GSQ-RCO? or both?

I continued to use Strata but something just feels off for this an large model, it is very fast but intelligence seemed even under 27B 3.8 UD Q4, it had a hard time coding a basic 3d game and a 2d platformer. Deepseek Flash V4 didn't have this huge and drop off at Q2 imo. Then I ran across this guys independent testing: lukesdevlab.com It shows that he needed many shots as well compared to the 27B. But then again Swift 1.5 27b is less smart, in his testing and mine. Even comparing it to the 27B swift 1.5 we can see it needs more compactions and more shots, for an far larger model. I would say just use the 27B or regular flash models, avoid Swift or Bottlecap if you are not so time sensitive since original delivers far better results in practice, in my own experience and the tests show it. Quants do play a part but I've used Q2 and Q3 quants before without this big of an issue, Strata and Swift are sacrificing too much intelligence for speed imo. TLDR: Depends on hardware (vram) Use Next flash strata if you need a very fast coding buddy who needs an exact detailed plan. Use 27B Q4 if you want the best results and have the time to wait (3x to 5x slower wallclock time) EDIT: he tested with Q1, I ran at Q3 but still I would say this doesn't even change the final conclusion for me! In actual use it hasn't been an intelligence jump, on the contrary to Q4 27B.