Was a heavy user of 1206 and currently under the assumption that 0205 is the replacement.
With that said, I think 1206 is the better version when it comes to creative writing. While both models have wonderful prose when prompted, 0205 falls behind due to worse repetition often entire paragraphs vs a few words, tighter parameters and overall less creativity, and more difficulties following a plot and character personality.
With the tight parameters its hard to increase temp for more creativity as it quickly falls into incoherence rather than creativity, so even when you swipe there is a lack of variety due to that.
While 1206 had its difficulties with abstract and complex storylines that was a bit easier to work around than the current difficulties with 0205 which have proven difficult to overcome. Overall, 1206, as a writer, was overall a better model. I picked it over offerings from the other major LLM companies.
Honestly, if 1206 became its own model this would be excellent.
Thanks a lot for the insight, its so rare to find people who use LLMs for actual creative writing since i know these companies mostly train their models to be good at coding and other things, creative writing is never their focus. Do you have any good tips for prompting when it comes to writing with the help of LLMs?
I’m also a writer who’s pretty devastated by the removal of 1206. It was the first LLM that actually managed to impress me with its prose (specifically for dialogue.) It could write crisp, strong dialogue with distinct character voices. It even had a good sense of comedy, which is an aspect most LLMs struggle with.
The new model is simply not capable of generating the same level of quality.
I’ve spent a lot of money yesterday on Openrouter API credits testing dozens of different models, open and closed weights alike, in the hopes of finding a replacement. Not one model has managed to continue 1206’s writing excerpts with a comparable level of quality. Not even close.
I’m taking a break from writing right now because I got so used to this new powerful tool that writing without it now feels off.
What temperature do you use for creative writing with Gemini? Personally, I’ve always valued the Claude models more for writing. Gemini seems to have a closed path when it comes to creating something with life. Maybe the developers aren’t taking that path seriously and are focusing on making the model create good code. As an amateur who has only written a few stories for the internet, I feel like Gemini is simply shackled when it comes to writing with a creative spark.
Well in terms of 1206, coupled with my specific prompting, I was able to use anywhere from 1.75/.94 to .90/.97 (temp/topP) Both of these still had great creativity and complete coherence, though I will admit 1.75 toed the line sometimes.
I agree that devs aren’t paying attention to creative writing and I understand that everyone wants coding and creative writing is a niche(sadly), that’s why I wish they’d just bring 1206 back for that purpose.
I was a big claude user as well. Opus was incredible and with some specific prompting, I felt 1206 was reaching for it. It couldn’t quite match it in the complexity department but compared to other LLMs it was my favorite.
Right now with 0205 pro, I’m messing with 1.49/.93 and some other variations 1.33/.94; 153/.90 But its just feeling like a worse model than 1206. Flash-lite gives me some of that same spark I got from 1206 though, def needs different parameters than 0205. start lower. I havent exactly dialed that model in but around 1.03/.97 should be a decent starting point.