@Fumbledew@FinalSpace We're the shipping not such a stinger I would have grabbed one too, I don't need it as I have standard midnight and colour on the way but it would have looked awesome on the shelf
@kapicode Setups I have had: 1x3090, 1x3080ti mobile 16gb modded to pcie and a CMP100-210 (v100mining card) and I had a batch of 10 of those cmps in the hopes they could be hacked to run well together, they couldn't, so a 32b was the best I ran at any usable speed
@IntCyberDigest Oh no what will I do, oh yeah just get a cloud server and setup my own for like £6 a month of hosting costs, or just use tor etc there's plenty of options
@haydendevs More than you can afford, even running heavy quantized you'd only get average performance on deepseek-v4-flash with like 256gb vram (like 3x rtx6000 pro at like £10k each plus the rest of the rig, ventilation etc, prob £35-40k all told then you have to power it)
@lauriewired So dig out some old hardware or head to an arcade with real vintage machines, many games from back then were hard though, they were designed either to bleed coins out of you at an arcade or last a long time on a home system as there just weren't many new games then
@starmexxx I don't know about a 235b, maybe if you're running at like Q2 but that's not a usable quant, you want Q4 min which wants usually about 1gb per 1b params. I believe the upcoming evo x3 with the 495+ will be available in up to 192gb even that isnt really enough
@humzaakhalid You're not replacing SOTA models with a 14b, if reducing bills from Claude level models is what you need start using deepseek-v4-pro for the less difficult work, it's absurdly cheap and way more capable than any 14b
@CooperZurad@mourginakis It's more prod soft that is public facing, modules get deprecated, APIs change etc so code needs updating to continue to work small private projects are very different to commercial stuff