Bijan Bowen compares Q4 and Q8 versions of Ornith 1.5 35B across coding, visual and agent-style tasks. The mixture-of-experts model activates only about three billion parameters at a time, making its useful results notable for local inference.
Bijan Bowen finds Q8 more consistent on demanding game generation, visual detail and quality-assurance passes. Q4 is faster and lighter, and it occasionally matches the larger quantization, but it is more likely to miss requirements or need repair on complex work.
Bijan Bowen concludes that Q4 is the sensible starting point for constrained hardware, while Q8 is worth the added memory and latency when output quality matters more than speed.
Watch the original on YouTube