M5 Ultra Mac Studio Review

Those are some incredible graphs, that leap in prompt processing going from M3 to M5.

Also: ~30 token/s on GLM 5.3-flash, locally. (That's roughly Opus 4.8-tier. I think).

/meta Here's a CSS filter that stops those nuisance chart animations,

    macstories.net##*:style(animation: none !important; transition: none !important)

A dense 27B doesn't really make sense for the Mac. A MoE makes way more sense when you have modest bandwidth but lots of memory.

Those RTX 5090 numbers are bad. You can get over 200 tps with ninfer using NVFP4 and MTP.

Thank you for this. I wish Apple focused their silicon design on improving the TTFT metrics but coming from an M3 Pro, it still looks laggard compared to Nvidia's TensorCores in the 5090.

Maybe Apple is an acquisition away from changing that balance.

The selling point of the M5 Ultra Mac Studio is that you can run much larger models that the 5090 can't without swapping. NVidia aggressively segments the market on VRAM for this reason. That's why a 5090 has an MSRP of ~$2k (but good luck getting one for less than $4k) while a 6000 Pro, which is basically a 5090 with 96GB of RAM has now soared beyond $15k where 3-6 months ago it was more like $10-11k. A 6000 Pro has the same memory bandwidth but slightly more CUDA units (IIRC ~24k vs ~21k).

This advantage won't be apparent with a 27B model. The 256GB MS can probably run the newer Flash models locally, something you can't do on a 5090.

I don't think we'll get a successor to the 5090 until late 2028, maybe even 2029. I'm basing this on the launch date of the 5000 series and that we haven't got a midcycle refresh yet. Rumor has it the chips are ready but the 3GB RAM modules are 3-4x the price of the 2GB modules used on the current cards.

Apple should see a Mac Studio major update in 2028. That might even force NVidia's hand. But it's really impossible to say what the state of the market will be 2-3 years from now. It may have completely crashed. I suspect not however.

The interesting thing will be when the bandwidth demands start forcing HBM memory onto these home/enthusiast solutions.

Not forgetting of course that an RTX5090 is what 600W+ ? And the Mac is probably half that at most ?

That's a dense model. Of course it will do worse.

Now try running that Qwen 3.8 Next model on the 5090 and tell me what TPS you get (hint: it's near 0 since it doesnt fit the 32GB VRAM on 5090 vs the 256 in OPs M5).

I assume those are non-batched. I think the M series GPU can do 4X to 8X depending on model quant, which means if you can batch queries you'll get almost 4X to 8X performance.

On my m5 max 27b model does 75tps on 256k ctx and starts at 80 on the 8k ctx when you add https://huggingface.co/collections/z-lab/dflash-2 to it. So yeah base might be 30tps (I used iq4) but mtp or dflash help a lot and should be used when checking what is useful and what is not for running models as it is not fare to judge without them.

The next Ultra, supposedly on deck in 2028:

> Apple's planned M7 Ultra chip is being designed to support up to 1.5 TB of unified memory and to push AI performance toward the class of Nvidia's Blackwell accelerators, according to a new Bloomberg report published by Mark Gurman...

Apple plans to release a base M6 chip this fall for entry-level Macs... a base M7 in the first half of 2027, M7 Pro and M7 Max at the end of 2027, and the M7 Ultra in 2028.

https://www.tomshardware.com/tech-industry/semiconductors/ap...

I really appreciate seeing these dense model numbers. For a large unified memory system though I expect that MoE numbers are what people are more interested in.

These numbers could and should get much better. As an example I can run Qwen3.8-27B-MXFP4 (W4A8) on 2x AMD R9700 that gets 260+ tokens/sec to start and slows down to ~110 tokens/sec over 128k context and can do the max 256k. These are for batch size 1 and throughput goes higher with batching. This is due to speculative decoding, efficient all-reduce inter-gpu compression, and custom GEMM kernels for the specific hardware. Note each R9700 only has 644 GB/s memory bandwidth.

A basic llama-bench on Qwen 3.8 27B UD-Q4_K_M gives pp512 3920 tok/s / tg128 81 tok/s on a 500W RTX PRO 6000 (should be similar speeds to a 5090, chip is basically the same, just less VRAM). With MTP3 this is 140 tok/s on mtp-bench.

This is with llama.cpp. You can of course use vLLM/SGLang well on these cards and they're even faster. On vLLM w/ NVIDIA/Qwen3.8-27B-NVFP4 baseline has a prefill of about 13,000 tok/s. The baseline tok/s is 72 tok/s, but at mtp7, it's 157 tok/s, and w/ dflash7 that goes up to 215 tok/s. On mtp-bench, DFlash2 gets a hair under 300 tok/s w/ the code_python prompt.

Yeah that can't be right, my M5 Max gets almost those speeds and certainly lot faster than what they're claiming the M3 ultra gets. Maybe they didn't have the model setup right or were running it with some unnecessarily high quant (>=8bit).

> This is great as a first look, but the author is not a developer, so we don't yet know whether a dev can be as productive with local models on M5 Mac Studio compared to a 20x subscription plan.

Local models are definitely not as productive as SOTA, sadly it's not close yet. I do think someday they will be "good enough" to use, but they aren't today. Even the SOTA models barely code well, with Opus 4.5 being the first, good coding model.

That being said, I think it's absolutely imperative that we keep pushing local model performance. We need to continue to advance technology there and ensure that the model labs don't do regulatory capture in the name of "safety" (or anything else).

Astra-Ultra? Even the largest open model to date (Kimi K3) is nowhere close to Astra level, and it will be quite slow even on the highest-spec M5 Ultra, with achievable speeds of about 0.5 tok/s at most due to having to stream weights from SSD (~13 GB/s on the highest storage capacity M5 Max machines so far). This is OK for doing simple Q&A in the background but it's far from a genuine coding experience. You'd have to test batching of multiple thinking streams in order to try and raise overall tok/s via layer-wise reuse of the streamed weights (and this is where the "Ultra" part sort of becomes relevant; Kimi series models have good support for agent swarms) but this would decrease single-session performance even further. It would only be usable for background jobs, though the hardware would then have a chance of paying for itself if it was fully used on a 24/7 basis.

Agreed.

Specially since one can pay half right now to OpenAI and sign a 12 year iron clad contract for uninterrupted service delivery of OpenAI Pro.

Yeah, anyone who thinks local AI is going to save them money is likely to be disappointed, at least if they want to run models that are even remotely capable.

Plenty of other reasons to get excited about local AI, but I don't think cost is one of them.

Hard to guess, it can go either way. If you will need to be in a syndicate to use non-sterilized models, that mac makes sense. But if there is mandatory registration of personal cyberarms, you risk going to mines once they check you purchases. You could try to play normie and pretend you simply wanted to show off, by keeping your actual work on external disk, but that leaves traces on system. Counting on someone in the Gap renting you gray iron works as long as you can swap credits. Still, this gear is tiny. Put it in your e-car, with uplink, and leave it at uncle's farm. Discreet.

Just commenting here because you're discussing hardware: I thought the test results from the SSD published in this article [1] were pretty interesting. Maybe that's old news though.

[1] https://www.macworld.com/article/3238319/mac-studio-m5-max-r...

512 option isnt worth it imo, you get severe slowdowns when weights are that large. 256 is the sweet spot, you can run large open weight models at decent speeds for full private inference.

5090 has 32GB VRAM.

It'd be silly to buy the 18k model to run a tiny model like Qwen 27B. You use models like GLM Flash and Qwen Next which won't fit on a single 5090.

Speed vs task-completion is a new conversation and a great point. Whereas the cost to compute doesn't exist in a vacuum, making mistakes costs less, is easier to maintain with granularity and a whole host of other factors when you own the lab.

What exactly is missing from macOS that makes you feel the need for Linux?

I get it on Windows systems, at least when someone wants to use Linux-type tooling. But macOS already supports pretty much all of that natively?

The Asahi Linux group goal should have been organizing themselves by taking out a license on Arm processor (probably too late with the present owners) and building a new Linux Distro to go with it instead of being a parasite on someone else’s existing hardware, if the three ex-engineers from Apple could do something like that why couldn’t they aim higher?

Why waste time spending years reverse engineering. Wouldn’t it have been easier to raise money for such an endeavor? I hope someone will come along in the future in kindergarten, junior high or high school who doesn’t know any better will try something like this. I’m too old. for such a journey.

This one is 2x m5 max, so ~1.2TB/sec.

I think M1 through M3 were compute bottlenecked in prompt processing (hence the very large gap between M3 and M5, in this page's benchmarks, that's not explained by memory bandwidth alone).

For generation speed in isolation, yes.

Another comment approximates at around USD$15k, so yeah, not zero.

I think the power itself will probably be more than you're paying in a subscription

i use linux a ton too, but still wonder what you want in a Linux server that macos as a BSD server does not have.

How are you measuring memory usage? top tells me that 45/48GB is "used", but Activity Monitor shows me that 24GB is cached files.

I'd be quite surprised if Mac OS alone needs more than 8GB, given that they sell the Neo with 8GB of RAM today.

Free RAM is wasted RAM.

Double?? They are nearly $100k.

You run Hermes and have a personal assistant.

You run uncensored local models where you can ask questions that would get denied by public providers, or questions that you prefer them not to know the intricate details (like your financial planning)

The solutions will come in the past, big iron, thought they were immune and not too long afterwards, personal computers took over and big iron was history well, the same thing is going to happen all over again.

Those companies building those big data centers are going to find out that they overspent. Because we’re not going back to the mainframe era, no matter how much OpenAI, Anthropic, Microsoft, Meta or Google would like to.

5 years of (200/month) tokens at that price, meanwhile an rtx 5090 pc is about half that… hmm

but i wonder how much these token costs are sustainable or not, it may be in the long term cheaper to have your own hardware if token costs go up (and hopefully hardware gets cheaper again)

The irony here is, thats hobby hardware. You spend 20k and can run slow hobby models that are barely capable of anything unsupervised.

Entry level serious hardware starts at 100k, and a bit better but still almost-useful grade is 200k (8x rtx pro, plus a nice epyc pairing). Thats the sort of thing a salaried expert lets their employer buy them for sort of serious work.

Anything really serious is well north of 1M - not including the housing and commercial grade mains connection. And at best that buys fast Kimi K3 or GLM.

The year of linux on the Desktop was 26 years ago for me.

For 99.99% of people, spending 15 grand on a Mac Studio just to run Qwen 3.8 locally is a non starter.

And nightmare fuel is just what they'll be selling at the UN this week, for this very reason.

Sam's address will probably be more riveting, imaginative, and terrifying than the last couple of Terminator screenplays. Legislators will lobby him to write the laws for them, and the ghost of Harlan Ellison will threaten to sue him.

I don't see how that math works? This is a $15k rig under benchmark and per the results it competes very acceptably against... one consumer GPU.

I really don't see who buys this, except people who want the Studio for some other reason. But nothing in the story says you want to fill racks with these instead of Blackwell or TPU parts; it's not even close.

> unless you have a bunch of money to throw at hobby projects.

there are lots of people with very expensive hobbies, see sailboat racing for example.

$5/hr = $3600/mo.

Unless you only need the AI available some of the time, $5/hr is pretty expensive. That's an RTX Pro twice a year.

If you're using it for discrete sessions of coding or something, that might make sense for you but if you're using it for an always-on assistant, that pricing kinda sucks.