I bet you do live under a rock making a 10m chat model. Only way to get something so impressive done without seeing the mamba-3 walk by and be like damnnnnn! Then go 3 steps back adding it. Us surface dwellers (at least this one) have 3 things going on and the only reason I'm working on something is because I thought of a better way.. And nothing gets done.
Complex-valued state update, implemented as a data dependent RoPE so it costs almost nothing. This is what gives it state tracking. Mamba-2 sits at chance on parity and bracketed modular arithmetic, Mamba-3 solves them. Has trapezoidal + complex-state changes and the half-state-size. But you surely have more support with mamba-2. May or may not be worth it..
https://goombalab.github.io/blog/2026/mamba3-part1/