ling 3-flash support

[smailiMohammad70]

any chance get a support in colibri ?
is like small and good model

[rofl0r]

is like small and good model

indeed. it's roughly on par with laguna S 2.1 both by size and capability.

[smailiMohammad70]

base on my exprience with smaller version leguna x2.1 worst than ornith35 on code agent but kat dev 2.5 was great model
if laguna s 2.1 is better i would try it

[rofl0r]

laguna s 2.1 worked pretty well in my testing, far better than qwen 3.6 35b and kat dev 2.5 - but i used the APEX compact quant ( Kwaipilot_KAT-Coder-V2.5-Dev-APEX-MTP-I-Compact.gguf ) and it performed pretty badly in my testing. also quickly regressing in t/s as context fills up. did you test ornith vs kat dev? which one is better?

[smailiMohammad70]

I used q6 https://huggingface.co/bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF on unsloth in code mode i got around 30t/s without mtp head and 92000 max context but code was solid
ornith35 is not bad but rethinking in middle response for example in first create alot folder later leave them and create a banch folder again put code on them and waste alot of max conext
my test was a massive asp.net core microservice project

[rofl0r]

i tried both qwen 3.6 and kat 35b versions to produce this : github.com/openwall/john/pull/6025 but both failed miserably. supposedly easy task: copy existing sha512 code and modify it for sha384. after repeated self-test failures and dozens of iterations trying to fix it, i pasted the diff produced by qwen into the free qwen 3.8 max chat page and asked it to find the bugs - but it failed too. lots of time wasted. laguna s 2.1 got it right on first try (that PR is from laguna's work).

[smailiMohammad70]

i am sure 120b somehow better than 35b
this week was wild so many model release or announced
dots3-note Preview look so good base on their bench

[Limalski]

One upvote for this. Ling is really good model and gives better answer than others of its size.