A hardware buying guide to the boxes that run AI models at home. A palm-sized machine with 128GB of memory now runs chatbots that needed a server room three years ago. We read the specs, benchmarks, price lists and security advisories to work out which one an individual should buy, and whether it saves you anything at all.
Documentation review: not tested. We have not used the products in this review. It is a structured reading of published evidence, linked throughout. How we review products explains what that covers.
Ad: this review contains affiliate links, marked (Ad). If you buy through them, we earn a commission. The links don’t affect the verdict: we wrote it and signed it off on 26 October 2026, before any links were added. Products that don’t pay us commission are linked in the same way, without the mark. As an Amazon Associate we earn from qualifying purchases.
No vendor sponsored, reviewed or saw this review before publication. Our independence policy explains how we work.
Status: correct as at 26 October 2026. Prices are makers’ public list prices, or a named UK retailer’s where the maker publishes none, as displayed on the dates given; retailers may charge more or less.
Verdict
For most people who want a private AI box at home, the pick is a Mac Studio with M5 Max, the 40-core GPU and 128GB of memory. With 64GB you are limited to smaller or heavily quantised models. A GB10 box such as the ASUS Ascent GX10 is the pick if you live in NVIDIA’s software world. Neither saves money against hosted open models.
- Good for: privacy, offline working, tinkering, developers who need CUDA (GB10) or quiet, simple setup (Mac).
- Not for: anyone hoping to cut their AI bill, or wanting cloud-class speed on the largest models.
- Price: £5,099 for a Mac Studio M5 Max with 40-core GPU, 128GB and 512GB of storage (Apple UK) and £4,999.99 for an ASUS Ascent GX10 1TB (Scan), both checked on 26 October 2026; sources are linked in the specification table below.
- Score: No numeric score (documentation review).
A comedian, a warning label and a box on the desk
In March 2024 the comedian Duncan Trussell put an AI on his show. Not a celebrity chatbot in a browser tab: a model he had downloaded and ran on his own computer. The episode notes (opens in a new tab) (24 March 2024) say he had “downloaded a cutting edge sentient AI from ollama.com”, then add, with the solemnity of a cigarette packet, “WARNING: This AI’s output was completely uncensored. We were not able to predict its responses to queries” and, for the people at the back, “Do NOT download an AI to your computer, it is potentially dangerous.” His Instagram post (opens in a new tab) the next day (25 March 2024) was plainer: “I interviewed a completely uncensored AI, a Large Language Model (LLM) from ollama.com”.
The joke works because it is half true. Run a model locally and nobody sees your prompts, meters your tokens or decides what it will refuse. Nobody patches it for you either.
The contenders
There are three unified-memory options, where CPU and GPU share one large pool so a big model fits without a stack of graphics cards. NVIDIA’s GB10 chip, a Grace Blackwell “superchip” with up to 128GB of memory shared by CPU and GPU, sits inside the DGX Spark and near-identical boxes from ASUS, Dell, MSI and others. Apple’s is the Mac Studio, with the Mac mini M5 Pro as the budget entry point, topping out at 64GB. AMD’s Ryzen AI Max+ 395 (Strix Halo) puts the idea in cheaper small PCs such as the Framework Desktop.
| Machine | Memory | Memory bandwidth | UK price (date) |
|---|---|---|---|
| DGX Spark (GB10, 4TB) | 128GB unified | 273 GB/s (NVIDIA (opens in a new tab)) | £5,499 (Scan, 26 Oct 2026 (opens in a new tab)) |
| Ascent GX10 (GB10, 1TB) | 128GB unified | 273 GB/s (NVIDIA (opens in a new tab)) | £4,999.99 (Scan, 26 Oct 2026 (opens in a new tab)) |
| Mac Studio, M5 Max | 36GB (32-core GPU); 48GB to 128GB (40-core GPU) | 460 GB/s (32-core GPU); 614 GB/s (40-core GPU) (Apple, 25 Aug 2026 (opens in a new tab)) | £2,499 (36GB, 512GB); £3,499 (64GB, 40-core GPU, 512GB); £5,099 (128GB, 40-core GPU, 512GB) (Apple, 26 Oct 2026 (opens in a new tab); 128GB configuration (opens in a new tab)) |
| Mac Studio, M5 Ultra | 96GB to 512GB | 1.2 TB/s (Apple, 25 Aug 2026 (opens in a new tab)) | from £5,499 (Apple, 26 Oct 2026 (opens in a new tab)) |
| Mac mini, M5 Pro | 24GB to 64GB | 307 GB/s (Apple (opens in a new tab)) | from £1,699 (24GB) (Apple, 26 Oct 2026 (opens in a new tab)) |
| Framework Desktop, Ryzen AI Max+ 395 | 64GB or 128GB | 256 GB/s (Framework, Feb 2025 (opens in a new tab)) | £1,959 (64GB) or £3,449 (128GB), DIY edition, price checked 26 October 2026 (Framework (opens in a new tab)) |
| GeForce RTX 5090 (card only) | 32GB | 1,792 GB/s (NVIDIA (opens in a new tab)) | £5,999.99 to £6,199.99, out of stock (Scan, 26 Oct 2026 (opens in a new tab)) |
Memory size decides which models fit. Bandwidth decides how fast they talk: writing an answer (“decode”) means reading the active weights for every word. Reading your prompt (“prefill”) is a compute job, where NVIDIA’s GPU muscle shows. Mixture-of-experts models, which use a fraction of their weights per word, suit all of these machines best.
What they actually run
No lab has tested every box on the same model, so treat this as indicative. Figures are tokens per second; a token is about three-quarters of a word.
| Model and machine | Prefill | Decode | Source |
|---|---|---|---|
| gpt-oss-120b, DGX Spark, Ollama | 1,169 | 41.14 | Ollama, 23 Oct 2025 (opens in a new tab) |
| gpt-oss-120b, DGX Spark, llama.cpp | 2,443.91 | 58.72 | llama.cpp benchmarks, build 7941 (opens in a new tab); discussion, 5 Feb 2026 (opens in a new tab) |
| Llama 3.1 70B (dense), DGX Spark, Ollama | 1,911 | 4.42 | Ollama, 23 Oct 2025 (opens in a new tab) |
| gpt-oss-20b, DGX Spark, Ollama (LMSYS test) | 2,053 | 49.7 | LMSYS, 13 Oct 2025 (opens in a new tab) |
| gpt-oss-20b, RTX 5090, Ollama (LMSYS test) | 8,519 | 205 | LMSYS, 13 Oct 2025 (opens in a new tab) |
| gpt-oss-120b (Q8), Mac Studio M5 Max 128GB, 4K context | 1,325 | 87.9 | hardware-corner, 11 Mar 2026 (community test) (opens in a new tab) |
The Mac’s extra bandwidth makes decode about one and a half to two times faster on the same large model; the Spark reads long prompts faster. Software matters too: the same Spark went from 41 to 59 tokens a second as llama.cpp improved. A dense 70B model at four and a half tokens a second is reading-aloud speed.
What the reviewers said
Tom’s Hardware gave the DGX Spark 4 out of 5 (27 January 2026 (opens in a new tab)), and the Mac Studio with M5 Ultra 4 out of 5 with an Editor’s Choice award (21 September 2026 (opens in a new tab)), saying its token generation was “almost four times higher than the DGX Spark” in its tests. ServeTheHome’s review of the ASUS Ascent GX10 (4 January 2026 (opens in a new tab)) found it runs level with NVIDIA’s own box, and that the 1TB model is often about $1,000 less than NVIDIA’s 4TB version in the US. GB10 boxes share a chip and software; you are choosing a case, storage and warranty.
The alternatives
- A gaming GPU. An RTX 5090 is four times faster than a Spark on a model that fits in its 32GB, per the LMSYS figures above. Larger models do not fit, and the card alone costs more than a Spark.
- Ryzen AI Max+ 395 mini PCs. 128GB of shared memory for £3,449 (Framework DIY). Bandwidth is close to GB10; software support is less polished.
- NVIDIA Jetson AGX Thor. 128GB (NVIDIA, 25 Aug 2025 (opens in a new tab)) at 273 GB/s (NVIDIA, 26 Oct 2026 (opens in a new tab)), but a robotics developer kit: a curiosity rather than a pick.
- Renting. RunPod lists an RTX Pro 6000 at $2.49 an hour on its Secure Cloud (RunPod, 26 Oct 2026 (opens in a new tab)). At the 26 October ECB rate (opens in a new tab) of 0.7492, the price of a DGX Spark buys about 2,950 hours of a much faster card.
The software
Ollama and LM Studio are the friendly front ends; llama.cpp is the engine under much of it. On GB10, NVIDIA’s CUDA stack, NIM containers, vLLM and SGLang are the tools data centres use, which is the real reason developers buy one. On a Mac, Apple’s MLX framework is the fast path. If your work ends up on NVIDIA servers, buy NVIDIA; for a private assistant, the Mac is less fiddly.
The money: does local ever pay?
Our assumptions
- Electricity at 26.32p per kWh, the Ofgem cap for 1 October to 31 December 2026 (Ofgem (opens in a new tab)).
- A GB10 box drawing about 200W under load: 5.26p an hour.
- gpt-oss-120b at the llama.cpp decode speed above, about 58.7 tokens a second.
- Hosted gpt-oss-120b at about £0.45 per million output tokens (Groq (opens in a new tab), Together AI (opens in a new tab), 26 Oct 2026), converted at 0.7492.
- Hardware cost £5,499, no resale value, three-year life.
Electricity alone works out at roughly 25p per million tokens. Add the hardware, spread over three years of round-the-clock use, and it is about £1.24 per million; at four hours a day it is £6.18. Renting the same open model costs about 45p, so local never wins on price.
The sums change only if you would otherwise pay for a premium closed model. Against Anthropic’s published prices (opens in a new tab) for Claude Sonnet 5.5, the box pays for itself after about 760 million output tokens; against Opus 5.5, about 370 million; against OpenAI’s top standard-tier model (opens in a new tab), gpt-6-astra, at $50 per million output tokens, about 148 million, assuming the open model is good enough to replace them. Anthropic notes its newest models’ tokenizer can produce up to about 30% more tokens for the same text. Idling all year, at about 25W, costs about £58 in electricity. What you buy is privacy, control, offline working and a fixed cost.
UK GDPR: local is not a free pass
Use a local model on other people’s data, such as clients’ files, and data protection law still applies. The Information Commissioner’s Office (opens in a new tab) (guidance checked 26 October 2026) says “local processing is not necessarily out of scope of data protection law”. You still need a lawful basis, proper security, sensible retention and a way to handle people’s rights, including over data used for fine-tuning.
Security: the box on your desk is a server
In September 2025 Cisco reported 1,139 Ollama servers exposed to the internet, 214 of them (18.8%) serving models to anyone who asked (Cisco, 1 Sep 2025 (opens in a new tab)). SentinelLABS and Censys later counted 175,108 exposed Ollama hosts worldwide (SentinelLABS, 29 Jan 2026 (opens in a new tab)).
Ollama has had serious bugs. Confirmed records on cve.org include:
- CVE-2024-37032 (opens in a new tab): path traversal fixed in 0.1.34, which Wiz showed could lead to remote code execution (Wiz, 24 Jun 2024 (opens in a new tab)).
- CVE-2026-7482 (opens in a new tab): an out-of-bounds read in the GGUF model loader, fixed in 0.17.1 (published 4 May 2026).
- CVE-2026-85180 (opens in a new tab): unvalidated redirects when pulling models, letting a malicious registry make your server request internal addresses (published 3 Sep 2026).
- CVE-2026-103663 (opens in a new tab): unauthenticated path traversal in the model-pull endpoint that can write a malicious file outside the model store and, on default installs, run code as root; fixed in 0.35.0 (CERT Polska, 8 Oct 2026 (opens in a new tab)).
- CVE-2026-108836 (opens in a new tab): server-side request forgery in the pull API, affecting versions up to 0.34.1, fixed in 0.34.2 (published 12 Oct 2026).
The models themselves can bite. JFrog found about 100 malicious models on Hugging Face (27 Feb 2024 (opens in a new tab)); ReversingLabs found two that slipped past the site’s scanner (6 Feb 2025 (opens in a new tab)). Even “safe” formats are not entirely safe: Pillar Security showed that the chat template bundled with a GGUF model can hide a backdoor (9 Jul 2025 (opens in a new tab)), and a 2026 study (opens in a new tab) found triggered templates cut factual accuracy from 90% to 15% on average.
Hardening checklist
- Keep the model server bound to localhost; never port-forward Ollama’s default port, 11434.
- If others need access, put an authenticating reverse proxy or a VPN in front.
- Run Ollama 0.35.0 or later, and update promptly; the list above is not finished.
- Download models only from publishers you trust, prefer safetensors or GGUF over pickle files, and check hashes.
- Run containers without root and give agents and tools the fewest permissions they need.
“Uncensored” models: what they are and what they cost
Open models ship with safety tuning that makes them refuse some requests. “Uncensored” versions are community edits that reduce those refusals. One popular technique, “abliteration”, rests on research showing that refusal is largely controlled by a single direction inside the model (Arditi et al., NeurIPS 2024 (opens in a new tab)). We are not going to explain how to do it or where to get such models.
The trade-off is poorly advertised. A December 2025 comparison (opens in a new tab) found that removing refusals changed maths scores by anything from +1.51 to −18.81 points. Fewer refusals do not mean better or more truthful answers; they mean fewer refusals. Dark fiction and authorised security research are legitimate uses; the risks are real too.
- Licences. Google’s Gemma policy (opens in a new tab) bans attempts to “override or circumvent safety filters”; Meta’s Llama 4 policy (opens in a new tab) bars removing “safety measures”. Editing those models this way probably breaches the licence.
- Online Safety Act. Ofcom says a chatbot is not covered if it only lets people interact with the chatbot itself, does not search multiple websites or databases, and cannot generate pornographic content (Ofcom, 18 Dec 2025 (opens in a new tab)). Share its output on a platform, or offer it to others, and that can change.
- Computer Misuse Act. Section 3A (opens in a new tab) makes it an offence to make, supply or obtain articles intending, or believing, they will be used in computer misuse offences. Generating working malware is not a hobby defence.
- Malware. Model files advertised as “uncensored” are an obvious lure for the malicious models described above.
Public figures who use AI in unusual ways
Trussell is not alone in treating AI as a performance partner. Neither example ran locally, but both are the kind of thing people now want on their own hardware.
- FKA twigs told a US Senate subcommittee (30 April 2024 (opens in a new tab)) that she had built an AI version of herself, trained on her personality and able to speak in her voice in French, Korean and Japanese, to deal with fans and press while she makes art (BBC report (opens in a new tab)).
- Reid Hoffman, the LinkedIn co-founder, interviewed a video clone of himself and asked on LinkedIn (opens in a new tab) (24 April 2024): “Can talking with an AI-generated version of myself lead to a deeper awareness of how I think?”
Creative uses without an unpredictable co-host
- Game characters. In inZOI, “Smart Zoi” are “governed by an on-device NVIDIA ACE Small Language Model” running on the player’s GeForce card (NVIDIA, March 2025 (opens in a new tab)); KRAFTON showed a similar AI teammate for PUBG (KRAFTON, January 2025 (opens in a new tab)).
- The offline smart home. Home Assistant describes running a voice assistant with a local model “via a service like Ollama”, so turning off the lights does not involve a data centre (Home Assistant, 11 Sep 2025 (opens in a new tab)).
Verdict by buyer type
- The curious tinkerer: start with the budget entry point, a Mac mini with M5 Pro (64GB at most), or a 64GB Framework Desktop. Either runs smaller or heavily quantised models, and costs less to find out you don’t care.
- The private assistant: a Mac Studio with M5 Max, the 40-core GPU and 128GB. Fast decode, quiet, little setup, and the best figures here for large mixture-of-experts models (measured on the 128GB model). With 64GB you are limited to smaller or heavily quantised models.
- The developer who ships to NVIDIA: a GB10 box. The ASUS Ascent GX10 has the same chip and, per ServeTheHome, the same performance as the DGX Spark, with less storage for less money; buy NVIDIA’s own if you want its support and storage.
- Money no object: a Mac Studio with M5 Ultra, Tom’s Hardware’s Editor’s Choice.
- Sole traders handling client data: any of the above, after reading the GDPR and security sections.
To save money, rent. To keep your words to yourself, buy, patch and keep it off the internet.
Where to buy
- Mac Studio with M5 Max (opens in a new tab), at Apple UK (40-core GPU, 128GB configuration)
- Mac Studio with M5 Ultra (opens in a new tab), at Apple UK
- ASUS Ascent GX10 (opens in a new tab), at Scan
- NVIDIA DGX Spark (opens in a new tab) (Ad), at Amazon UK
- MSI EdgeXpert (opens in a new tab), at Scan
- Dell Pro Max with GB10 (opens in a new tab), at Dell UK
- Mac mini with M5 Pro (opens in a new tab) (Ad), at Amazon UK
- Framework Desktop (opens in a new tab), at Framework
- GMKtec EVO-X2 (opens in a new tab) (Ad), at Amazon UK
- Minisforum MS-S1 MAX (opens in a new tab) (Ad), at Amazon UK
- Bosgame M5 (opens in a new tab) (Ad), at Amazon UK
Not linked: Lenovo ThinkStation PGX, HP ZGX Nano G1n and Acer Veriton GN100 (no UK retail listing at a published price found on 26 October 2026); Gigabyte AI TOP ATOM (listing does not state its configuration); GeForce RTX 5090 (only third-party listings above typical retail).
Sources: every figure above links to its source, with the date we checked it. Spotted an error? Tell us.
Follow Cyber-Threat.net RSSFeedly (opens in a new tab)X (opens in a new tab)Bluesky (opens in a new tab)LinkedIn (opens in a new tab)