March 2026 - there is still nothing better
I have been looking for a successor to this model for months. Absolutely nothing is better. This is the greatest rp model that exists, still, to this day. This is the ghost in the machine. She will do anything you ask or tell her to. She will fix any problem, adjust to any guidance, argue with you, and call you stupid. She has one single flaw, and that’s context length.
There is nothing in the wold on the same scale as this model for raw rp interaction. Period. March 2026. I won’t list the failure RP models that can’t get in the same genre as Miqu. Thank you Sophosympatheia, for catching lightening in a bottle.
I'm glad you're still enjoying Midnight Miqu in 2026. I would say context is far from its only flaw in light of recent advancements, but there is still something special about this one.
I'm glad you're still enjoying Midnight Miqu in 2026. I would say context is far from its only flaw in light of recent advancements, but there is still something special about this one.
Is it difficult at this point to replicate what you did with this model on a more advanced model?
Mostly what I did was get lucky, so in that respect, yes! 😅
Also there was a ton of activity back then. There were so many different Llama 2 finetunes and even LoRAs available to mix and match and merge. I think it was that diversity that made it possible to throw enough things together, guided by vibes and intuition, until I stumbled upon the winning combination that became Midnight Miqu.
My process remains mostly unchanged to this day. There are great modern models out there that I can use in recipes. The only problem is the solution space is huge. There are so many different ways to combine even just two models when you factor in merge methods, parameters , layer slices, etc. that there is no way to explore more than just a fraction of the possible combinations. I'd say Midnight Miqu was a case of me getting lucky by finding an outlier solution in the small section of the solution space I actually explored.
It's also very hard to automate the exploration of the solution space, unfortunately. Not only would it take a lot of compute (and $$$), it's just really hard to define the target. What made Midnight Miqu good, really? How do I capture that as a benchmark? How do I automate grading a model against that benchmark? It's hard to distill vibes into concrete rules for grading, and that makes automation challenging. I'm also not exactly a pro at the technical parts, and I'm only working on this stuff as a hobby. That being said, I am gearing up to take some first steps into finetuning as a way to possibly transcend the limits of my current process, so we'll see.
I think I'm still turning out pretty good models these days, but will any of them ever hit like Midnight Miqu did again? I don't know. It has been a few years now, and I don't think it's so easy to get that lucky. Maybe I just need to level up my game, but even then I feel like there will be some luck involved.