Close Menu
    Facebook X (Twitter) Instagram
    partnerships@markellismedia.uk
    YouTube X (Twitter) Instagram LinkedIn TikTok Newsletter
    Mark Ellis Reviews
    • Home
    • Reviews
    • Tech News
    • Opinion
    • Buying Guides
    • Tech Tips & Guides
    • Advertise
    • Newsletter
    • Dead Pixel Club
    • Apple
    • iPhone
    • Android
    • MacBook Neo
    • iPhone 17e
    • M5 MacBook Air
    • M4 iPad Air
    • M5 MacBook Pro
    • Best Phones
    • Best Headphones
    • Best Apple Watch
    Mark Ellis Reviews
    Home » Blog » Reviews » M5 Ultra Mac Studio Review: Goodbye Claude?
    Reviews

    M5 Ultra Mac Studio Review: Goodbye Claude?

    Apple's most powerful desktop meets local AI - is the Mac Studio the future of private, cost-free LLM workflows?
    Mark EllisBy Mark EllisUpdated:October 8, 20261 Comment12 Mins Read

    We may earn compensation from reviewed products. Learn about our editorial policies.

    90%
    90%

    The M5 Ultra Mac Studio is overkill for most, but for businesses serious about private, cost-free AI workflows, it's a genuinely compelling piece of capital expenditure. The ability to run multiple frontier-scale LLMs locally - and access them remotely - is transformative.

    The Good
    1. Runs multiple large LLMs simultaneously without breaking a sweat
    2. Fully private, local AI processing with no cloud costs
    3. Remote access via Tailscale works seamlessly across devices
    4. Offline vibe coding and image generation both impressively capable
    The Not So Good
    1. Eye-watering price at top-end configurations
    2. Local models can't replace frontier LLMs for up-to-date knowledge
    3. Setup requires meaningful technical effort
    • 9
    • User Ratings (0 Votes)
      0

    Apple Mac Studio (M5 Ultra, 256GB)

    £12,299

    ✅ Pros:
    – Runs multiple large LLMs simultaneously with ease
    – Fully local AI processing – no cloud, no token costs
    – Compact form factor with impressive port selection
    – Remote LLM access via Tailscale is seamless
    ❌ Cons:
    – Extremely expensive at top-end configurations
    – Local LLMs can’t match frontier models for up-to-date knowledge
    – Setup requires some technical know-how

    See Amazon Price

    It’s a cliché, but we really are living in interesting times.

    Slightly scary times, admittedly.

    I am, of course, referring to the incessant march of AI. Right now, it feels like an out-of-control nuclear missile heading toward a future most of us only ever expected to see in movies.

    What that future is, is anyone’s guess. I remain open-minded but extremely cognisant that AI is a genuine threat to our livelihoods, way of living, and, possibly, our lives.

    The latter feels like an absolutely insane way to begin a review of the new Mac Studio, but bear with me.

    We use AI in my business. It’s for the boring, mundane, and number-heavy stuff that I’d rather the team didn’t waste time on. Our adoption of AI for these tasks, reports, and planning has given us time back to invest in more creative, exciting things. But we’re relying entirely on Claude and, therefore, Anthropic, for all of this. That means sending important data into the cloud and allowing it to rest on servers we have zero control over.

    This is why I was so keen to try out the new Mac Studio. I’ve never purchased a Mac Studio because it always felt like overkill for what we do. That mindset is shifting, though, because, clearly, the latest super-powerful Macs are designed, among many other things, to run AI locally. That’s what this review is focused on.

    Why I’m Doing This

    I appreciate that in order to obtain access to the most up-to-date, powerful, and capable large language models (LLMs), I need to pay Anthropic for access to its frontier wizardry.

    I’m also conscious that most of the research I do for this business relies on querying things that happened yesterday.

    A local LLM doesn’t help with either of those requirements (maybe the latter, to a degree, but it still needs an internet connection for that). What it does help with, however, is the ability to run a lot of the querying we do behind the scenes, which doesn’t require frontier LLMs or the latest news.

    If you weren’t aware, you can download a vast number of LLMs from many different providers and leave them running, ready for prompting, on your computer. If it’s powerful enough. Which the Mac Studio absolutely is – particularly the one Apple sent me to review.

    It doesn’t cost a penny to download, load, and query those LLMs. They’re yours to use, for free.

    The Mac Studio? Not quite so ‘free’.

    What Is a Local LLM?

    A local LLM (Large Language Model) is an AI model that runs entirely on your own hardware – no cloud, no subscription fees, and no data leaving your device. Tools like Ollama make it straightforward to download and run these models on a capable Mac.

    The Mac Studio Specs

    The Mac Studio I was sent to review is about as far from the base spec as you can imagine.

    It has the top-end M5 Ultra chip, which contains a 36-core CPU, 80-core GPU, and 32-core neural engine. The memory has been upgraded to an almost comical 256GB (a £4,000 option). The only thing that hasn’t been maxed out is the storage, which on this review unit is 4TB (it goes up to 16TB if you have a spare £7,500 handy).

    The price when you place this configuration into your basket is a trouser-destroying £12,299. Which you could of course use instead to buy a second-hand car.

    At least it has a couple of USB-A ports and a front-facing SD card slot. You don’t get that on a Vauxhall Astra.

    This might seem astronomically expensive for a computer. And for mere mortals like myself, it is. But if you’re running a business which invests heavily in AI, it’s nothing more than a smart piece of capital expenditure. As always, Apple makes these universe-bending specs available for a reason.

    The Software Stack

    So, what do we do with all of that power?

    Firstly, we need some software. Most of what I’m about to link to is completely free, so knock yourself out and give it a try yourself.

    Ollama is best thought of as the engine. It’s the thing into which the local LLMs (I’ll get to them in a second) are loaded. You don’t ‘use’ it – Ollama just sits in the background doing its thing.

    Msty Studio is the ChatGPT-style interface that can be used to query those local LLMs. I won’t be using it in this particular test because I want to focus on how one can remotely access those LLMs, and for that, we need something else – Open WebUI. This provides the exact same chat-style functionality as Msty Studio, but it runs in a web browser and can be accessed by multiple people, wherever they happen to be.

    To make Open WebUI accessible for the team, I had to install something called Tailscale, which creates a private network into which I can add specific computers and mobile devices. That enables anyone to access the Open WebUI chat interface and query the LLMs sitting on the Mac Studio.

    Lastly, we have MacWhisper for video-to-text transcription, Draw Things (I’ll let you guess what that one does), and Claude Code running locally.

    But on to those LLMs.

    Three Models Loaded at Once

    For this test, I downloaded three LLMs:

    Qwen 3.6 35B-A3B (35 billion parameters, 20GB)

    Llama 3.3 70B (70 billion parameters, 43GB)

    gpt-oss 120B (120 billions parameters, 65GB)

    In addition, I downloaded Parakeet V3 for MacWhisper and FLUX.1 (schnell) for Draw Things. Both ensured that I could run those two apps offline and still have access to powerful LLM processing on the Mac Studio.

    I have since published the video version of this test on YouTube and, perhaps expectedly, have been hit with a whole bunch of “why are you using that model?”, “those models are out of date, mate”, “I can’t believe you didn’t use Qwen 3.9 LBAAABN Z4.56 Ver 4.5 rev 5421B.1” comments. Despite offering the caveat at the start of the video that I had no idea what I was doing.

    Regardless, the point of this test isn’t to max out the Mac Studio (I don’t know how to). It’s to demonstrate what’s possible, and you can use your imagination to assess just how far this could go.

    Curious how Apple Silicon handles real-world workloads? Check out our M4 MacBook Air review for more.

    Remote LLM

    In my M6 Mac mini review, I ran everything locally on that brilliant little computer. That’s great if you work alone or have no reason to access your LLM setup remotely. For the Mac Studio, I wanted to try a more relevant test that involves accessing the LLM from any computer or smartphone.

    As mentioned, Tailscale makes this possible, and it’s a self-contained network that isn’t accessible by anyone apart from those whom I let in. That is, immediately, reassuring. Remember, we’re also querying local LLMs here, so at no point do my or my team’s prompts get fed into a server I have zero control over.

    All we have to do, once connected to Tailscale (it essentially sets up a private VPN connection on the device or computer you’re using), is open a specific web address, log in, and start a chat within Open WebUI. Just like you would with Claude or ChatGPT.

    Those prompts get fed into any one of the three local LLMs running within Ollama on the Mac Studio. And because this particular Mac Studio has an industrial-sized volume of onboard unified memory, we can keep all three running at once (more on that in a second). You just choose the model (or models) to query.

    I also loaded a bunch of business data onto the Mac Studio and full transcripts from my last 50 YouTube videos. This enabled me to not only query the LLMs’ knowledge, but also dive into specific business data – and combine the two. It worked, brilliantly.

    I urge you to try this. Mainly because the level of calm that came over me when I reminded myself that I was in full control of what was happening – and that what was happening was still bonkers for a child of the 80s – was extremely lovely. No token costs. No sending my business data anywhere but the internal storage on that Mac Studio.

    Is this the future?

    Tailscale at a Glance

    Tailscale creates a private, encrypted VPN-style network between your devices. Once set up, anyone you invite can securely access the Open WebUI interface on your Mac Studio from anywhere in the world – no public server required.

    Three Brains, One Question

    In an attempt to push the Mac Studio a little, I realised that I could query all three LLMs at once. I’m not sure why I’d ever need to do this, but it sounded like a cool test.

    This still involves Open WebUI; you just select all three models, type in your prompt, and hit the Enter key. Within seconds, I had three answers – one from each model- within the same chat window. Once again, I could do this from any device I’d connected to my Tailscale network. And, once again, all of the processing was taking place on the Mac Studio.

    Using all three models at once and, indeed, having them loaded constantly into memory, resulted in just a few gigabytes of remaining unified memory being available on that world-crushing Mac Studio. The memory pressure also shot up while the LLMs were doing their thing, and there was a fair amount of swap file stuff going on. Understandably, I’d guess.

    This is where you start to think about the folk who could actually make use of all that power. It’s relatively easy to max out 256GB of unified memory if you’re a heavy local LLM user, and the headroom provided by that monstrously powerful M5 Ultra chip means multiple users can do their thing without hindering one another. Again, without any form of spend on frontier models.

    Exciting, right?

    It gets better, too.

    Vibe Coding with the Internet Off

    I have a tricky relationship with vibe coding. We use it all the time in my business and have built some incredible tools as a result of simply chatting to AI.

    Despite this, I’m under no illusion that the stuff we’ve built is far from being product-ready. It works (most of the time) brilliantly for internal stuff, but if you’re serious about developing apps, you still need a human brain to oversee it. I spent years in a former role working with some very talented developers, and I firmly believe that their skills are more important in the age of AI than ever before.

    However, if you do want to develop software for your business, or just for fun, you may not be aware that you can run Claude Code locally. It’s really simple to get running, too – just a bit of work in Terminal, and giving it access to one of your locally running LLMs is all you need to do.

    I did this on the Mac Studio, unplugged the internet, and told Claude Code to create a timer app for me that could run in a browser. It did it in less than a minute, and it worked perfectly. Because of course it did – this is the simplest example of local vibe coding.

    Just think of the possibilities, though. This kind of self-contained development, arguably, doesn’t require access to a frontier model. You can build, experiment, and create things without spending a penny on tokens and, once again, for any business or individual that can attach this to productivity gains or revenue generation, computers like the Mac Studio become a seriously tempting prospect.

    Want to see how Siri’s on-device AI stacks up? We put it through its paces – read our Siri AI review for more.

    The Image Race

    As mentioned earlier, I installed Draw Things on the M5 Max Studio. I gave it a relatively complicated prompt that involved a bicycle, a puddle, some oranges, and 35mm film.

    Once again, this image generation was all running locally. And it did it in around 38 seconds.

    I’ll leave that there, because I have zero interest in AI doing creative work for me.

    Verdict: Who Is This For?

    Once again, I must reiterate that this review wasn’t focused on maxing out the Mac Studio. That requires more time than I had, and a brain far bigger than mine.

    What it has done, for me at least, is convince me that I do need a Mac Studio for our business. The ability it will give us to access local LLMs remotely and do work without relying on Anthropic is genuinely exciting.

    The question is: how much power do you need?

    I’m just glad Apple gives us so many options these days. It’s a very good time to be in the market for a new Mac.

    Shop the Apple Mac Studio (M5 Ultra) below!

    See Amazon Price
    Apple Apple Silicon Mac Studio
    Mark Ellis
    • Website
    • Facebook
    • X (Twitter)
    • Instagram

    Mark Ellis has been the Chief Reviewer and the driving force behind MarkEllisReviews.com since its inception in 2020 (the guy the website’s named after!). As the founder of the site, Mark has built a reputation for delivering relatable tech reviews and advice for normal people. Now Mark primarily reviews the biggest tech releases from Apple to Android and attends events across the globe to keep his finger on the pulse of upcoming developments. He also makes videos for his YouTube channel with an audience of over 260,000 subscribers.

    Related Posts

    iPhone Duo Review: A Big Surprise

    9 November 5, 2026 Reviews

    iPad mini 8 Review: I Didn’t Expect This

    8 November 3, 2026 Reviews

    Linkind Smart Permanent Outdoor Lights ColorBlend Review: Christmas Lights You Only Install Once

    November 6, 2026 Reviews

    1 Comment

    1. Placarunt on October 15, 2026 8:30 pm

      Running AI models locally sounds like a great option for businesses that want more privacy and fewer ongoing API costs. The biggest challenge is finding the right balance between hardware costs, performance, and access to up-to-date information. It’s similar to checking vehicle details through RUNT por placa in Colombia as well as you can update your information of personally at portal, where having the right information available makes a real difference. I think local AI will become increasingly useful for repetitive tasks, while cloud models will still have an advantage for complex research and real-time information. Would you consider investing £12,299 in a Mac Studio, or is a cloud-based AI subscription still the better option?

      Reply

    Leave A Reply Cancel Reply

    Survey
    RECENT
    • iPhone Duo Review: A Big Surprise November 5, 2026
    • iPad mini 8 Review: I Didn’t Expect This November 3, 2026
    • Linkind Smart Permanent Outdoor Lights ColorBlend Review: Christmas Lights You Only Install Once November 3, 2026
    • Why I’m Excited About Apple’s First Touchscreen MacBook October 30, 2026
    • Google Fitbit Edge Review: Brilliant Simplicity for £160 October 28, 2026
    Get my free weekly newsletter!

    Access my free community, plus weekly behind-the-scenes content, exclusive perks, and cool finds—all in one email.

    About

    Mark Ellis has been writing about and reviewing tech since 2019, amassing an audience of well over 1 million regular viewers, readers and listeners. Free from jargon and complicated benchmarks, Mark and the team deliver reliable guidance and updates about the latest technology for normal people.

    X (Twitter) Instagram YouTube LinkedIn TikTok
    • Home
    • Reviews
    • Tech News
    • Opinion
    • Buying Guides
    • Contact
    • About
    • Advertise
    • Editorial Guidelines
    • Privacy Policy
    © 2026 Ellis Media Group.

    Type above and press Enter to search. Press Esc to cancel.