The good thing with most NNs is that you can go with fp32 almost always (I put almost because I'm not 100% sure that this covers all the cases) for almost 0 performance degradation. For Parakeet I'm 100% sure because I used it yesterday and it was using just 1GB.

Why do you use GLM 5.3 tho? Unless you have an instance running "free" 24/7 you could use a local Gemma model. It should be good enough, no? Or maybe with openrouter it's cheap enough for such simple requests to be considered free?