This looks super promising — linear decoding complexity while improving on longer context lengths up to 1M! Super interested to see how this does when scaled up.

→ View original post on X — @jiquanngiam