Gryphe

/

Pantheon-Reasoning-27B

Like 31

I'm curious about this too, but I'm also curious if other folks have taken their 3.6 tunes and applied them to 3.8, and how that's played out.

Anecdotally, 3.8 may have less general knowledge, but better programming or agentic skill, so, it may not actually apply to the creative writing or roleplay use cases. (But it might!)

I'm curious about this too, but I'm also curious if other folks have taken their 3.6 tunes and applied them to 3.8, and how that's played out.

FrenzyBiscuit from ReadyArt has released several finetunes of Qwen 3.8, using the same datasets as previous versions based on Qwen 3.6 (e.g., DarkScarlett 1.0 & DarkScarlett 2.0, Serenity 1.0 & DarkScarlett 2.0), and in my opinion, the 3.8 versions are significantly better in terms of intelligence (instructions following, remembering what’s happening in the story, etc.), while the writing remains just as good, which makes RPing with them nicer.

From what I've seen on their Discord, other users are also positive about 3.8.

Alright, alright, since there's such a clear demand I'll see what I can do! Honestly, I've gotten quite curious myself now, lol.

Just make sure you submit it over at caliperbench.com after - seems to be an interesting RP evaluation site.

If anyone knows a good base to tune it on, I'm open to suggestions. I know base Qwen is a bit of a pain when it comes to censoring, but there's so many choices it's too easy to lose sight of it all!

If anyone knows a good base to tune it on, I'm open to suggestions.

You can check out this comparison of uncensored variants by Abliterlitics: https://abliterlitics.dev/models/qwen38-27b/

orcarouter/Qwen3.8-27B-Uncensored seems to be the best bet.

Thanks, just what I needed! I'll cook something up with the Orca base, then.

I can't believe I'm going to say this, but, on that site HoodedPirate gave, it actually does look like DavidAU's "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU" model actually scores really well. And no, that's not an exaggeration of the name, that's the real name.

Benchmarks are one thing, but user reports are conflicting, including complaints about losing track of things compared to official release, as well as premature end of generation.

Imo, just to be safe, I'd suggest working with the official release. Maybe whip a style tune first, you seem to have been on a roll with those, and did one on a MoE 3.6 qwen as well. It would be a shame to put blind faith into an abliterated base only for it to turn out to make the result semi-borked. But you're the expert, not me, so it's up to you at the end of the day, just my 2 cents