@leploutos@lxucan Mimo 2.5 is available free on OpenCode. It’s _fine_ for frontend but needs hand-holding sometimes, but the vision is nice. As long as you have examples elsewhere it can reference it does the job.
2.5 Pro is about Sonnet level I’d say, but no vision.
@leploutos I run Mimo 2.5 on my Nanobot. Extremely cheap little assistant that also supports vision!
I use it to count calories, handy to throw it a screenshot of my food and it does the rest.
Just gotta throw like $5 a month at it, and I can always use 2.5 Pro on OpenCode if I ever need
@weswinder@AnthonyDo Remember that dumb time for like 2 weeks right before agents when we had to pass every file to them in the chat that we wanted them to modify and look at?
That sucked.
Thank goodness for tools.
@gmoneyNFT M5 has some beefy gains over M4 on the neural stuff. Each GPU now has a neural accelerator for up to 4x faster compute. Memory bandwidth is also 30% faster than M4 (better TPS)
Worth the wait for September/October when the M5 Mac Mini/Studio drops imho
I’m eyeballing a 128GB
@TheAhmadOsman But for real though, floppy disks used to only hold a couple MB. We’re back in floppy disk era.
Eventually, running these models will be as mundane as a 2TB SD card. Expensive, maybe, but doable.
@ryanflorence@chantastic +1 for anxiety/depresssion meds suck 💯 You get to choose between either weight gain, insomnia, or stomach pain the rest of your life with the hope that _maybe_ it will slightly take the edge off (plot twist, it won't).
I vote we give them free Fable to develop new alternatives
@ryanflorence@kentcdodds@ChrisEsplin I can't do the teardown videos anymore. Feels like a jumpscare the whole time waiting for the part where he knifes the edge of the device. I can't do it 🫨
Nails on a chalkboard I tell ya
@AnuragShar74342@tannerlinsley > Omg the goat itself
Oh, yeah, that’s normal, you said the “TanStack” word. It summons the goat and his team. And no, it never stops being cool feeling like your voice is heard.
THAT level of support is another justification in itself to choose TanStack in my opinion 🤷♂️
@weswinder@theo@teortaxesTex Too much time and compute chasing marginally better models, not enough reinventing the core architecture for better efficiency.
The JEPA "world model" idea is interesting - thinking in abstract vectors instead of tokens. Excited to see both cost and performance gains from that.
@weswinder@theo@teortaxesTex Definitely shaken the Chinese model industry to compete for price.
Too bad U.S. labs are only incentivized to be the most intelligent and not the cheapest.
They gotta get our money while they can, eventually on-device compute will be as mundane as 1TB fitting on a flash drive.
@weswinder@theo@teortaxesTex Nope, I am wrong! I had it in my mind for some reason that both Flash and Pro were 75% off until the end of May but only Pro got extended indefinitely.
I stand corrected - Flash is just dirt cheap with no discount.
Well crap. How am I gonna break the news to my nanobot now 🥲
@weswinder@theo@teortaxesTex I am torn between using Flash for free on OpenCode, or getting my dirt cheap money’s worth thru DeepSeek directly until the end of May 😂 it’s a good problem to have.
My nanobot assistant will be getting a brain boost to V4 Pro once Flash loses it’s discount. Just needs vision
@weswinder@theo@teortaxesTex Gotta disagree on this one. V4 Pro is VERY capable thru OpenCode harness. Even Flash is enough most days. DeepSeek has become my daily driver. I think if you gave it a try for a day you would be impressed.
It likes to use subagents so it’s a little slow, but worth it for price