Views
No views yet
[SYSTEM_PROMPT])
[SYSTEM_PROMPT])After using stupid shit like 26B gemma4 finetunes for weeks. going back to roc xl is nice
I am using it currently. With 12b models like cydonia i had problems with the AI not understanding some nuances, even with thinking. This model with thinking pushes way above it's weight in my opinion, feels more like a 24b to 32b model. At least judging from my current brief experience
wow, it writes really well! It's really close to a Cydonia. And, I mean v4.3+ Cydonia, not before. If I was sacrilegious, I'd say close to v4zr. I'll continue using it, it has really a huge potential. But yes, it punches really hard for its size. When I have some time, I'll try to set up reasoning <thinking>, I can put some Q4 all into VRAM with a bit of space left. Really curious to see how it goes!
The model prose is way better than 12B - the former often writes in a specific way, and 16B is far more flexible.
I'd like to echo what other people have been saying about Rocinante XL 16B v1's prose, it's actually really solid and often beats out the last Magidonia/Skyfall versions I tried in that regard.
So far it's goated for me. Using meth
it really is like night and day when using metharm vs using teken its not even funny lol
Honestly I'm not using dry or xtc the rest is pretty similar so far I prefer it over most models in the 20B~ range
Yeah iv been bouncing around .8-1.1 and overall it's pretty good! I think logical inconsistencies are the weak point (I still think 12b Rivermind Lux mogs most models even to this day in that front) but the rest is really good it's shocking it's a 16B. Much better than Snowpiercer and I think it holds up well to something like Cydonia.
Been benchmarking Rocinante XL 16B for long-form RP over the last few days and honestly I’m pretty impressed.What stands out the most isn’t the prose quality (which is already very good), but the character agency and emotional continuity. The model consistently takes initiative, closes scenes when it makes sense, introduces new directions naturally, and generally feels more like it’s roleplaying a character than just reacting to user input.One thing I particularly liked: the model occasionally refuses to follow nonsensical narrative pressure from the user. For example, I had a scene where the character was half asleep and I kept trying to continue the conversation. Instead of forcing more dialogue, Rocinante simply advanced the scene to the next morning. That kind of narrative judgment is surprisingly rare.Character consistency has also been excellent. In my tests, emotional conflicts evolved naturally over time instead of resetting every few messages. Motivations felt persistent and consequences actually mattered.The main weakness I found is context degradation. Up to around 16k context everything feels fantastic. Around 20k I start noticing drift. Past that, repetition, weird callbacks, and occasional nonsense become increasingly common. By 30k+ the quality drop becomes very noticeable. This isn’t unique to Rocinante, but I figured it was worth mentioning.Interestingly, I got much better long-term results by using aggressive narrative summaries (~2k tokens of relationship state, character state, emotional state, and open threads) plus a small recent tail, rather than feeding huge amounts of raw chat history.Overall I’d easily rank it among the strongest RP-focused models I’ve tested. It genuinely surprised me several times, which is probably the highest compliment I can give a roleplay model.Great work.
its the only real improvement the old nemo had in a year
config-v1a