I also find that the 35B A3B in Q4 outdoes Ternary Bonsai 27B. Mind you, because I am not that interested in long-context, fire-and-forget tasks, I think the f

I also find that the 35B A3B in Q4 outdoes Ternary Bonsai 27B. Mind you, because I am not that interested in long-context, fire-and-forget tasks, I think the faster MoE model has more value generally.