So how to find the mamaba for you,
- Standard mamba model; honestly pretty great but is it worth the hype? Absolutely have u seen how much transformers get? Narcs
- Hybrid models; alr now we're talking honestly mamba-1.58 version is a wet dream if u don't care about precision as much, linear attention (keeps ur tokens and tokens per second [t/s] in check), u can also use zamba2 I feel it's legendary,
- Ultra Low ram users like me; best bet is sub-1B if pure ram,and u want good speed,and not want ur device to go bye-bye and kamakaze
- Power users who r rich 🤑; the world is ur to take it recommend liquid AI, it's multimodal as well, if u feel extra lengthy and girthy, try learning to graft or auto distill from a larger model like deepseek R1 so this babe actually can do autonomous tasks, the reason is precision, SSMs are notorious for batshit precision, tho mamba3 does solve most of it
- Overall good family, very open source, and barely any incest
Top comments (0)