M5 Ultra Mac Studio review: 256 GB of unified memory makes local AI agents viable
- MacStories tested the M5 Ultra Mac Studio with 256 GB of unified memory against an M3 Ultra with 512 GB and a desktop PC with an RTX 5090, and calls it a dream machine for local AI agents.
- The reviewer made Qwen3.8-Flash-Next running locally his default model in both Open Minis for iOS and Hermes Agent, and runs local subagents inside the Codex app orchestrated by GPT-6 Astra.
- The review's speed chart for Qwen3.8 27B has the M5 Ultra at 48 tokens/s at an 8K prompt down to 24 at 256K, versus 59 down to 44 on the RTX 5090, which cannot hold a 256K context at all.
- An RTX 5090 still has higher memory bandwidth, but the Mac's unified memory runs models such as GLM Flash and Qwen Next that do not fit in the 5090's 32 GB of VRAM.
- Commenters priced the tested config near $12,299 and a 2 TB/512 GB unit above $15,000, against roughly five years of a $200/month cloud subscription; the 512 GB option ships in October.
Hacker News opinions
$12,299 as configured. That's five years of a $200/month token plan, and an RTX 5090 PC is about half the price. If token prices climb and hardware gets cheaper again, owning it wins out.
The numbers I actually want are buried in a chart near the bottom. Qwen3.8 27B at 8K/64K/128K/256K: 59/51/44/n-a on the 5090, 48/39/32/24 on the M5 Ultra, 31/23.5/20/15 on the M3 Ultra.
30 tokens/s on GLM 5.3-flash locally, roughly Opus 4.8 tier. Also, here's a CSS filter to stop those chart animations: macstories.net *:style(animation: none !important).
A dense 27B is a bad benchmark for a Mac. When bandwidth is modest and memory is huge, a MoE makes way more sense.
Those 5090 numbers are bad. You can get over 200 tps with ninfer using NVFP4 and MTP.
The whole selling point is running models the 5090 can't hold without swapping. Nvidia segments hard on VRAM: a 6000 Pro is basically a 5090 with 96 GB, and it's gone past $15k now after sitting at $10-11k three months ago.
Don't forget the 5090 pulls 600W+. The Mac Studio is probably half that at most.
On my M5 Max, 27B does 75 tps at 256k context with dflash-2 and starts at 80 on 8K. The base numbers undersell it if you don't use MTP or dflash.
The next Ultra is supposedly 2028: an M7 Ultra designed for up to 1.5 TB of unified memory, aiming at Blackwell-class AI performance, per Gurman.
The 512 GB option lands in October and probably adds another $4-6k, so a 2 TB Ultra is north of $15k. That's about 12 years of OpenAI Pro.
Anyone who thinks local AI saves money is going to be disappointed, at least if you want remotely capable models. There are plenty of other reasons to be excited, cost just isn't one of them.
512 GB isn't worth it, you get severe slowdowns once the weights get that large. 256 is the sweet spot, you can run big open weight models at decent speed for fully private inference.
The machine he tested is $18k as configured. I did not expect that to make a 5090 look like a good deal.
A 5090 has 32 GB of VRAM. Buying an $18k box to run a tiny 27B model is silly. You'd run GLM Flash or Qwen Next, which don't fit on a single 5090.
Yes, the hardware is great. But I can't take anyone seriously who says macOS isn't a hindrance. If it had proper Linux support I'd care, otherwise I'd only touch it as an LLM server.
What exactly is missing from macOS that makes you need Linux? It already supports most of that tooling natively.
The author isn't a developer, so we still don't know if a dev can be as productive on local models as with a 20x subscription. If programmers get an Astra-Ultra-like experience all week on this hardware, that's the real tipping point.