Post

Ultramarine-1-flash — a GLM-5.3-Flash fine tune that outperforms Fable at 1/60th the cost open weights, open harness, open training data, open training code, so open.. www.paperinstruments.com/blog/ultrama...

A.V.

@slckl.bsky.social

Very interesting. Unfortunately, the coding takes a hit, even compared to GLM-5.3-Flash.

mr. TIM

@timkellogg.me

my $0.02: who tf cares? we got lots of coding models, what about everything else? we really don’t need another coding model, and if this was the best, it would be superseded in a week anyway good on them for finding a focus

15/10/26, 11:38 am1 repost10 likes

A.V.

@slckl.bsky.social

Just a shame that they couldn't keep at least the baseline performance. Coding is the one domain where AI clearly works, other knowledge work domains remain a bit fuzzier.

Martin #509

@martin509.bsky.social

I'd be curious to see if better knowledge and so on translates to better soft skills at coding tasks, e.g. interpreting requirements correctly, better structure, etc

OκULTRA XP🧿, writer for The Economist

@0kultra.bsky.social

I say someone could finetune in other direction and make a "coding better subspecies"