Date: June 2nd, 2026 2:37 AM Author: Burgundy useless brakes nowag
lol there's nothing better than Qwen3.6 27b that will fit on that card. You won't even use half your VRAM, but anything better requires more than 96gb. You either need 48gb or 256gb, and nothing in between really counts
On one server I have a 5090, a 5080, a 5070 ti, and a 5060 ti, but I'm only using the 5090 and 5080 to run my Hermes Agent right now, because I only need 48gb to run Qwen3.6 27b at Q8. I have two 3090s sitting in another server and I use them to do OCR, translation, image gen and shit like that. I just have a bunch of LXC containers running Ollama on the 3090s, and my Hermes Agent connects to whatever Ollama container it needs for a given task. I used to run a separate model for coding tasks, but Qwen3.6 does that well enough now. I only paid about $700 apiece for the 3090s plus another $100 to get them re-pasted by a shop, now they go for like $1200 on ebay with 6-year old paste lol
Can you even buy M4 Ultras? I've been looking at Mac Studios for a few weeks, and they only let you spec it with a M3 Ultra or M4 Pro. And they won't let you add more than 96gb of RAM (which to be fair is plenty).
Anyway the Ultra models are never getting cheaper. The M5 is the same design as the M4 and M3, just made using a newer process at the TSMC foundry. You get insane memory bandwidth with all of them, it's just a matter of how many GPU cores you have to crunch on it. A M2 Ultra probably beats M5 Pro for AI just because it has more GPU cores.
GPT/claude said Macs can run larger models but would be so slow that it would practically be unusable. Like waiting an hour for a response or 8 hours to do a full brief draft
Right. I think the ideal Mac has to be the Studio with an Ultra -series and 64gb VRAM. That lets you use 48gb comfortably with room to spare. I wouldn't plan on using more than 48gb with a Mac and I can only see 96gb being useful if you want to have two models running concurrently. What I like about my GPU spread is the ease of serving up multiple LLMs concurrently l. My OCR -> vector embedding workflow is pretty ridic, and I have a server that translates media from any language and produces an English transcript.