Getting a 33B model onto a 12GB GPU sounds like the hard part. It’s actually deciding what you can afford to compromise. That’s what stood out from this Apodex 1.1 test on MiniMax H3. The brief kept changing: first speed and quality, then audio became non-negotiable, then an existing setup was handed over for review. The result changed with it. Pruned Q4 over full Q4, no TeaCache, SageAttention only if it actually proves useful, and a much less optimistic 8–15 minutes for 8 seconds of video. Sometimes the useful answer is knowing which optimization to leave out.