if i can do good FFI abstractions in rust im sure opus 5.5 could handle it

man it's not a trick question i just want to know what sort of capabilities would cause people to move off the frontier subs and onto their own local hardware

clarification because people are misinterpreting: - the model will prefill and run as fast as qwen3.8 27b does on your hardware - this isn't a trick question about never being able to use another model i'm just curious as to what sort of capability would cause people to unsub

i think the tweet implies usability

i wonder if you could partner with educational institutes for this sort of thing. it's a pretty mutual relationship? they give you known-good pre-AI papers (with student permission, of course) and in return they get some discount on pangram seats

like i was using opus 4.8*** havent used it in months

yeah i’ve been using 1017 like i’ve been using opus 4.8. it’s been wonderful. it’s annoying because deepseek serve the model at 250tk/s but it makes so many bad guesses that it feels slower than opus 4.8 ever was despite being 4x faster

did you get the two mixed up? 😭

yeah agreed. i’d love a local opus 5.5 but something like opus 4.8 in every way (knowledge, speed, thinking duration etc) would work fine for me

assume the frontier hasn’t moved from today. you still can’t really do reverse engineering with the frontier. would you accept a local opus 4.8 equivalent for the task if it had no safeguards? what about gpt-5.5? that’s sort of what i’m asking; how good is good enough

assume a hypothetical LLM can fit into whatever hardware configuration you have. what is the minimum equivalent capabilities you’d cancel your frontier subs for, to use the model locally permanently?

huge

can u put this somewhere so i can play with it

its crazy how they chose to release this checkpoint to begin with given it was bad even for the time. qwen3.6 27b beat it on release and qwen3.8 27b beat it harder (bearing in mind how much cheaper these models are to host)

i think it takes genuine skill to make a 218B model perform so terribly whilst still being coherent

we WILL be promoing

i thought this said mogged