All roughly the same price, for a consumer who has won the lottery:
- 1x 512 GB Mac Studio with the 4 TB disk storage option.
- 4x Nvidia GB10s (total 512 GB RAM) each connected with a 100 GiB/s link to a Mikrotik CRS504-4XQ-IN.
- 3x Ryzen AI Max+ 495 192GB systems (total 576 GB RAM), connected via a 40 GiB/s link directly to each other.
Has any individual/SMB purchased this yet? Aware the ROI isn't there for most individuals that don't heavily use LLM for coding or multi agentic workflows -- but for those that it does make sense this is ~$68k cheaper than the next alternative making this the most realistic option for consumer grade access to frontier models.
Specifically interested if anyone has ran GLM-5.3 (753B MoE) in 4-bit on a 512GB Mac Studio. Looking for real world usage numbers and/or any benchmarks you might have performed.
They haven't shipped yet (people who purchased the moment it became available have delivery dates near the end of the month). That said you can estimate it's speed by looking at the 256GB version for models with similar numbers of active parameters (GLM 5.3 has 40B active parameters). So for 4-bit you could compare to what people are getting Qwen3.8 27B at 8-bit (which is around 2000 tok/s prefill and 70-80tok/s token generation [1]). It might be a little higher or a little lower due to the different architecture and kernel optimizations.
GLM-5.3 would be bottlenecked by the M5's GPU, here. I don't have any hard numbers, but Apple Silicon is known for weak prefill performance on larger models (yes, even MoEs).
Good to know. Private GLM-5.3 is really the one use-case I can imagine for a system like this - It does very well with cybersecurity tasks, which are either blocked for most people by current providers, or on their way to getting blocked based on the regulatory/political trends. Such a capability could become very valuable when supply becomes artificially limited, although I expect the Chinese labs will be the last hold-outs. Imagine not being able to do security hardening on your own tech because government-imposed guardrails decide your legitimate prompts appear "risky". In such a world, what would an unconstrained cyber-capable model be worth?
Both. Apple Silicon could have the best of both worlds if it supported dGPU/eGPU tech, but Apple depreciated that about a decade ago. So you're stuck with fast RAM and a slow GPU.
What are you comparing this to, and for how much, and what sort of benchmark numbers is it getting? I'd love to run better cheaper hardware. I'll gladly pay you a reasonable finders fee.
Get a personal computer with 48-64 GB RAM (or maybe even 32 GB). Whenever your model usages go beyond that – there's those APIs/subscriptions just for that. You are golden for most (really most) of the work.
But if your use-cases need extremely heavy-duty local models then just open your bank vaults and get the one you need.
Can't really answer that without knowing your use case. What is the largest model you need to run? What are you going to do with it? What tps do you need?
For just trying out local models and running occasional queries a 16GB mac will run an 8B-20B dense quantized model just fine.
If you are running at the scale of OpenAI then a data center full of blackwells will have the best ROI.
Cheapest expensive? I'd say the M5 Pro Mini with 64GB RAM. That's certainly not cheap, but not ludicrous, and fairly capable. An M6 mini with 32GB would be the budget option and quite limited, you'd probably want to replace it after say 4 years.
To do significantly better than that with a Studio, say M5 Max with 128GB would cost twice as much again. It would probably be more than twice as performant and much more capable, but you'd be getting into very big bucks that IMHO would only be justifiable for paying pro work.
I recently bought the M5 Pro Mini with 64 GB of RAM, and it feels excellent. It matches the Max in CPU performance while running Qwen 3.8 27b. Although it is slow, I can complete tasks at what I would call Opus 4.6–4.8 levels. What surprises me is using Claude Code on it. The speed of building with that harness and testing iOS apps is impressive… full end-to-end control of Xcode through to TestFlight. This is a major upgrade from my old M1 Pro MBP.
Buy one of the 128gb m4 max when people sell them to buy m5 ultras. It runs fine on the 27b models and qwen3.8 can actually run tools and perform tasks well. I use it for private agents, voice processing (no elevenlabs criminal pricing) and some photo manipulation though it's not very efficient at this task.
Running inference on LLMs that can fit in 512GB. For example, you would be able to run either GLM-5.3-Flash or DeepSeek V4.1 Flash 1017 on this, both of which are very respectable models. It's up to you, of course, if ultimate privacy is worth $25k.
All roughly the same price, for a consumer who has won the lottery: - 1x 512 GB Mac Studio with the 4 TB disk storage option. - 4x Nvidia GB10s (total 512 GB RAM) each connected with a 100 GiB/s link to a Mikrotik CRS504-4XQ-IN. - 3x Ryzen AI Max+ 495 192GB systems (total 576 GB RAM), connected via a 40 GiB/s link directly to each other.
Your AI Max+ 495 would be hobbled by ~300gb/s bus no? So muuuuch slow inference
Your AI Max+ 495 would be bottlenecked by ~300gb/s no?
What is the 40GiB link ?
Probably USB4.
Well USB4/TB4 is 40 Gigabit/s, so 8 times slower than what they are expecting.
Does anybody actually expect bandwidth to be expressed in bytes, rather than bits?
Dupe/Discussion (67 points, 1 week ago, 82 comments) https://news.ycombinator.com/item?id=50277166
When the page first loads, it shows a different price for ~1sec. It quickly overwrites that and displays the much larger, eye watering figure.
1st load ref: https://i.postimg.cc/YC5V5yZH/prerewrite.png
That's the 96GB ram price.
what the f is that sketchy site doing?
Loading the page with default parameters, probably from cache, then applying the specified parameters
The state of the web when every site that is just text and images, for some reason uses javascript
Has any individual/SMB purchased this yet? Aware the ROI isn't there for most individuals that don't heavily use LLM for coding or multi agentic workflows -- but for those that it does make sense this is ~$68k cheaper than the next alternative making this the most realistic option for consumer grade access to frontier models.
Specifically interested if anyone has ran GLM-5.3 (753B MoE) in 4-bit on a 512GB Mac Studio. Looking for real world usage numbers and/or any benchmarks you might have performed.
Thanks!
They haven't shipped yet (people who purchased the moment it became available have delivery dates near the end of the month). That said you can estimate it's speed by looking at the 256GB version for models with similar numbers of active parameters (GLM 5.3 has 40B active parameters). So for 4-bit you could compare to what people are getting Qwen3.8 27B at 8-bit (which is around 2000 tok/s prefill and 70-80tok/s token generation [1]). It might be a little higher or a little lower due to the different architecture and kernel optimizations.
[1]: https://omlx.ai/benchmarks/performance?model=qwen3.8+27b&chi...
Crescent island will be cheaper and might ship earlier. It could also be faster.
GLM-5.3 would be bottlenecked by the M5's GPU, here. I don't have any hard numbers, but Apple Silicon is known for weak prefill performance on larger models (yes, even MoEs).
Good to know. Private GLM-5.3 is really the one use-case I can imagine for a system like this - It does very well with cybersecurity tasks, which are either blocked for most people by current providers, or on their way to getting blocked based on the regulatory/political trends. Such a capability could become very valuable when supply becomes artificially limited, although I expect the Chinese labs will be the last hold-outs. Imagine not being able to do security hardening on your own tech because government-imposed guardrails decide your legitimate prompts appear "risky". In such a world, what would an unconstrained cyber-capable model be worth?
I think 2xGB10 might be a better option but don't quote me on that.
is this a hardware or software limitation?
Both. Apple Silicon could have the best of both worlds if it supported dGPU/eGPU tech, but Apple depreciated that about a decade ago. So you're stuck with fast RAM and a slow GPU.
The elephant in the room: how much of those 512 Gb is usable by the GPU?
75% by default, ~90% with manual overrides.
apple always charged ridiculous numbers for ram and disk space.
What are you comparing this to, and for how much, and what sort of benchmark numbers is it getting? I'd love to run better cheaper hardware. I'll gladly pay you a reasonable finders fee.
You can get it up to 29k with max storage.
By comparison, the Macintosh IIfx with 4 MB of memory and a 160 MB hard disk was about $28k in today's dollars.
https://en.wikipedia.org/wiki/Macintosh_IIfx#Models
So, effectively just as worth it for those who need it.
It was a 10k starting price in the UK IIRC, there were two in our maths department!
I remember reading about the Apple Lisa when it was announced. Launched January 19, 1983, with 1 MB RAM: $9995
https://en.wikipedia.org/wiki/Apple_Lisa
Ship date is 3/5/27 for the 512GB model.
Oddly enough, Apple Upgrade isn't an option.
5/3/27 if you are in the UK.
5/3/27 if you're in one of the ~150 countries that use date first
27/5/3 if you're in one of the ~15 countries that use year first
5/3/27 If you're in the 3 countries that use month first
Guys let's all be reasonable and write the date normally:
2003-05-27
Thanks.
I usually write it 2026-Mar-03 in email these days because you never know which locale people are in.
ISO is the way to go.
Will one be enough, is my only question?
To be safe, just buy four and you'd be able to run the current biggest open weight models.
- gonna ask you HN smart folks a deep question
- one one hand, the 8GB RAM Mac mini with M6 as cheap as it maybe is an absolute waste of money when it comes to running LLMs
- on the other hand the most expensive machine is right here
- what is the cheapest most expensive machine you can buy capable of providing you the best ROI for running LLMs locally?
Get a personal computer with 48-64 GB RAM (or maybe even 32 GB). Whenever your model usages go beyond that – there's those APIs/subscriptions just for that. You are golden for most (really most) of the work.
But if your use-cases need extremely heavy-duty local models then just open your bank vaults and get the one you need.
Can't really answer that without knowing your use case. What is the largest model you need to run? What are you going to do with it? What tps do you need?
For just trying out local models and running occasional queries a 16GB mac will run an 8B-20B dense quantized model just fine.
If you are running at the scale of OpenAI then a data center full of blackwells will have the best ROI.
the use case is getting the best bang for your back per dollar
take an 8 year life cycle for all machines and think of the cheapest expensive machine that you can buy which ll give you a decent llm to run
Cheapest expensive? I'd say the M5 Pro Mini with 64GB RAM. That's certainly not cheap, but not ludicrous, and fairly capable. An M6 mini with 32GB would be the budget option and quite limited, you'd probably want to replace it after say 4 years.
To do significantly better than that with a Studio, say M5 Max with 128GB would cost twice as much again. It would probably be more than twice as performant and much more capable, but you'd be getting into very big bucks that IMHO would only be justifiable for paying pro work.
I recently bought the M5 Pro Mini with 64 GB of RAM, and it feels excellent. It matches the Max in CPU performance while running Qwen 3.8 27b. Although it is slow, I can complete tasks at what I would call Opus 4.6–4.8 levels. What surprises me is using Claude Code on it. The speed of building with that harness and testing iOS apps is impressive… full end-to-end control of Xcode through to TestFlight. This is a major upgrade from my old M1 Pro MBP.
Buy one of the 128gb m4 max when people sell them to buy m5 ultras. It runs fine on the 27b models and qwen3.8 can actually run tools and perform tasks well. I use it for private agents, voice processing (no elevenlabs criminal pricing) and some photo manipulation though it's not very efficient at this task.
Probably 64GB MacBook Pro and run a 27B model on it comfortably.
I hate to sound like Sam Altman, but is there actually any point in having 512 GB of RAM other than just pure tech hoarding?
Running inference on LLMs that can fit in 512GB. For example, you would be able to run either GLM-5.3-Flash or DeepSeek V4.1 Flash 1017 on this, both of which are very respectable models. It's up to you, of course, if ultimate privacy is worth $25k.
... Or even 640K
512 ‘ought to be enough
[flagged]