Buyer's guide · updated October 12, 2026
Local AI computers, compared
Bring the brain home.
Startups are betting that people will pay $1,799 to $3,499 for a box that runs AI on the desk instead of in someone else's data center. They join NVIDIA, AMD, Apple and a crowd of mini-PC makers. This is the local side of the local-vs-central fight Underclassic tracks every day. Here is every box we could verify, what it costs, and what it can actually run.
Disclosure: Underclassic may earn a commission when you buy through some links on this page. It doesn't change what we recommend or the price you pay.
The short list
Two numbers decide almost everything. Memory size sets the biggest model that loads at all. Memory bandwidth sets how fast it answers. Start with the model you want to run, then buy the cheapest box that holds it.
| Box | Price | Memory | Biggest model, comfortably | Availability | Best for |
|---|---|---|---|---|---|
| Mac mini | From $899 | 16 GB to 64 GB unified | About 14B at 16 GB; about 30B at 64 GB (est.) | In stock at Apple | Mac households who want to try local AI with Ollama or LM Studio before buying a dedicated box |
| Oki Home | $1,799 | 16 GB GDDR7 (RTX 5060 Ti) + 32 GB RAM | The 27B model it ships with | Batch 1 sold out (ships mid-December); Batch 2 reservations ship Q1 2027, US first | Households that want photos, files and an always-on assistant in one private box |
| Mac Studio | From $2,499 | 36 GB to 256 GB unified (614 GB/s Max, 1.2 TB/s Ultra) | 120B-class MoE on a 128 GB M5 Max; 200B-plus on a 256 GB Ultra | Shipping; long lead times reported on high-memory Ultra builds | Anyone who chats or codes with one big model all day and wants the fastest replies |
| Ghost Core | $3,499 | 24 GB GDDR7 (RTX PRO 4000 Blackwell) + 64 GB DDR5 | 30B class (ships with Qwen 3.8-27B and Gemma 4-31B) | Batch 1 sold out (Ghost's CEO, Oct 5) and ships October 31; batch 2 interest list open | People who want a set-and-forget box running personal agents 24/7, driven from a phone |
| GMKtec EVO-X2 | $3,649.99 | 128 GB unified LPDDR5X (256 GB/s) | 120B-class MoE (gpt-oss-120b); dense 70B runs at about 4 tok/s | In stock direct from GMKtec | Tinkerers who want 128 GB for big mixture-of-experts models at the lowest price, on Windows or Linux |
| AMD Ryzen AI Halo | $3,999.99 | 128 GB unified LPDDR5X | AMD says up to 200B parameters | Micro Center, in-store pickup only | Developers who want a supported, Spark-sized box on x86 with Windows or Linux |
| GB10 partner boxes | From $4,999 (64 GB) | 64 GB or 128 GB unified (GB10) | Same as DGX Spark | 128 GB patchy; 64 GB from October 23 | Teams that want the Spark with a corporate warranty, bigger SSDs or a familiar vendor |
| Surface RTX Spark Dev Box | $5,999.99 | 128 GB unified (NVIDIA RTX Spark, 20-core CPU, 6,144 CUDA cores) | Over 120B parameters (Microsoft) | Pre-order now; ships in November, US only | Developers who want CUDA and 128 GB on a Windows desk, with dev tools set up out of the box |
| Framework Desktop | From $6,799 (192 GB) | Up to 192 GB unified (Ryzen AI Max+ PRO 495, 273 GB/s) | DeepSeek V4.1 Flash at Q2 on 192 GB (Framework's own pick) | 192 GB batch 1 on pre-order for November; 128 GB out of stock | Linux users who want standard Mini-ITX parts and a path to clustering |
| 192 GB Ryzen AI Max+ PRO 495 boxes | From $6,799 (192 GB) | 192 GB unified LPDDR5X (273 GB/s; up to 160 GB for the GPU) | Up to 300B parameters (AMD) | EVO-X5 Pro shipping; MS-S1 MAX-P495 on pre-order, shipping from mid-October | Tinkerers who want room for 200B–300B mixture-of-experts models in a mini PC, on Windows or Linux |
| NVIDIA DGX Spark | $6,950 (128 GB) | 64 GB or 128 GB unified (GB10, 273 GB/s) | Up to 200B parameters on 128 GB; about 100B on 64 GB (NVIDIA) | 128 GB scarce; 64 GB on sale October 23 via partners | Developers who need CUDA, fine-tuning, or several agents served at once |
| tinybox | From $11,000 | 64 GB VRAM (4× Radeon 9070 XT) + 32 GB RAM; up to 192 GB VRAM with two NVIDIA RTX 6000 cards | 70B class split across four GPUs (est.) | New tinybox made to order, ships in 1–8 weeks; red v2 and green v2 out of stock | Small labs and serious hobbyists with a 1,700 W circuit and room for a 42 lb box |
The boxes, cheapest to dearest (by entry price)
Prices are US list prices from each maker's store on October 12, 2026 unless noted, and memory shortages are moving them monthly. “Buy” goes to the maker's own page. Each box ends with “What reviewers say”: written reviews, YouTube reviews and posts from X, all linking out. X excerpts are quoted exactly, and we don't embed players or X widgets. Reviews checked October 12, 2026.
The DIY baseline
Mac mini
Apple
The cheapest way into Apple's unified memory. The base M6 model is fine for 7B to 14B models. The M5 Pro configurations go to 64 GB, which is enough for 30B-class models with room for context. It's quiet and sips power. When TechCrunch interviewed Ghost's founder, this is the machine it asked him to beat.
What reviewers say
Written reviews
Video reviews
From X
Taedio affectus @TaedioaffectusOwner
“…Same box, same harness (Hermes): 75 tok/s on a Gemma 26B MoE, 11.7 on a 31B dense. Bandwidth is the wall.”
Manmohan @ManmohanBuildsOwner
“My short-chat winner lost on a long document.
Mac mini M6 32GB local LLM benchmark.…”
The family memory box
Oki Home
Oki Applications (Y Combinator)
Part home server, part assistant. Oki runs a 27-billion-parameter model on an NVIDIA GeForce RTX 5060 Ti with 16 GB of GDDR7, next to 32 GB of system memory, and keeps your files on a swappable credit-card-sized “Memchip” that scales from 2 TB to 16 TB. There is 10 Gb Ethernet and a Linux environment for tinkerers. Oki's FAQ is upfront that remote access, connected services and optional cloud models can send data off the box.
$1,799
Batch 2, 2 TB Memchip included; $59 refundable reservation; $2,299 after Batch 2
- Memory
- 16 GB GDDR7 (RTX 5060 Ti) + 32 GB RAM
- Biggest model, comfortably
- The 27B model it ships with
- Availability
- Batch 1 sold out (ships mid-December); Batch 2 reservations ship Q1 2027, US first
- Best for
- Households that want photos, files and an always-on assistant in one private box
What reviewers say Pre-release
Not shipping yet. These are launch coverage and posts from the company, not hands-on reviews.
Launch coverage
From X
Luofei @Luofei_ChenFounder
“Say hello to Oki Home, your Personal AI Computer.
We're building a future of limitless, personalized intelligence for every consumer, while keeping all your data in your hands.…”
Luofei @Luofei_ChenFounder
“A dedicated 4 vCPUs + 8gb ram on services like fly costs ~$132/mo. Pair that with api calls, $10-50/mo/user, cloud storage, you're easily looking at $200/mo/user in costs.…”
Fastest for one person
Mac Studio
Apple
Memory bandwidth is what makes a local model feel fast, and the Mac Studio has the most of it here. Tom's Hardware measured the M5 Ultra at almost four times the DGX Spark's token rate on Qwen 3.8-27B. The catch is software: MLX and llama.cpp cover inference well, but most fine-tuning code still assumes NVIDIA.
From $2,499
M5 Max; 128 GB M5 Max is about $5,099; M5 Ultra from $5,499
- Memory
- 36 GB to 256 GB unified (614 GB/s Max, 1.2 TB/s Ultra)
- Biggest model, comfortably
- 120B-class MoE on a 128 GB M5 Max; 200B-plus on a 256 GB Ultra
- Availability
- Shipping; long lead times reported on high-memory Ultra builds
- Best for
- Anyone who chats or codes with one big model all day and wants the fastest replies
What reviewers say
Written reviews
Video reviews
From X
Abhishek Yadav @yabhishekhdBenchmark recap
“2 Nvidia DGX Spark vs Apple M5 Ultra Mac Studio, both with 256GB memory, tested by Alex Ziskind, Mac wins in writing speed while Sparks are up to 3x faster in reading long prompts…”
Mladen Lotar @lotar_vkOwner
“Just got 20+ tps on Qwen 3.8 Flash Next on M5 Max with 128 GB!…”
Vlad @gitstatus001Owner
“Tried TensorFold and increased token/sec x2 on my Mac Studio M5 Max (compared to same model via Bionic)…”
The agent appliance
Ghost Core
Ghost
A screenless “brain in a box” built around an NVIDIA RTX PRO 4000 Blackwell SFF card (24 GB of ECC GDDR7, 432 GB/s), a six-core Ryzen 5 7600, 64 GB of DDR5 and a 1 TB SSD. It arrives with Qwen 3.8-Next, Qwen 3.8-27B, Gemma 4-31B and Muse-Glimmer-30B preloaded. Ghost came out of stealth on October 5 with an $11 million seed round led by Andreessen Horowitz; its CEO, Zain Javaid, is 19. He told TechCrunch the box keeps working even if Ghost shuts down.
$3,499
no subscription; 30-day returns
- Memory
- 24 GB GDDR7 (RTX PRO 4000 Blackwell) + 64 GB DDR5
- Biggest model, comfortably
- 30B class (ships with Qwen 3.8-27B and Gemma 4-31B)
- Availability
- Batch 1 sold out (Ghost's CEO, Oct 5) and ships October 31; batch 2 interest list open
- Best for
- People who want a set-and-forget box running personal agents 24/7, driven from a phone
What reviewers say Pre-release
Not shipping yet. These are launch coverage and posts from the company, not hands-on reviews.
Launch coverage
Video reviews
From X
Zain Javaid @zainmfjGhost CEO
“Batch 1 of Core has sold out!
We're excited to start shipping units out in the next few weeks.…”
Zain Javaid @zainmfjGhost CEO
“Introducing Core, The Personal Al Computer
The most powerful technology humanity has built should belong to you…”
Most model per dollar
GMKtec EVO-X2
GMKtec
The best-known AMD Ryzen AI Max+ 395 (“Strix Halo”) mini PC. ServeTheHome got Llama 3.3 70B running entirely on the integrated GPU at 3.7–3.8 tokens per second and Qwen3-32B at about 10.5, which tells you where it shines: big sparse models, not big dense ones. Memory prices have pushed it far above its original $2,000.
$3,649.99
128 GB / 2 TB; 128 GB / 1 TB is $3,499.99; 64 GB / 1 TB is $2,199.99
- Memory
- 128 GB unified LPDDR5X (256 GB/s)
- Biggest model, comfortably
- 120B-class MoE (gpt-oss-120b); dense 70B runs at about 4 tok/s
- Availability
- In stock direct from GMKtec
- Best for
- Tinkerers who want 128 GB for big mixture-of-experts models at the lowest price, on Windows or Linux
What reviewers say
Written reviews
- ServeTheHome: GMKtec EVO-X2 Review an AMD Ryzen AI Max 395 Powerhouse
- Tom's Hardware: GMKtec Evo X2 AI Mini PC review: Compact Strix Halo Powerhouse
Video reviews
From X
Jose Romero @hijoseromeroOwner
“Four Qwen 3.8 Flash-Next quants ran on my GMKtec EVO-X2, a 128 GB Strix Halo PC. My result on this setup, not a typical single-GPU claim.…”
Luka @l3d3kaOwner
“was thinking abt returning my evo-x2 64 but that post literally saved it, now it’s running the IQ3XXS quand faster than it was running qwen 3.8 27B”
The Spark rival that boots Windows
AMD Ryzen AI Halo
AMD (sold by Micro Center)
AMD's own developer box: the same Ryzen AI Max+ 395 as the EVO-X2 in a Spark-like 150 mm square chassis, with 10 GbE, a 120 W power supply and preinstalled AMD developer tools. StorageReview called it the only box in the Spark's class that boots Windows 11 or Linux on a full x86 platform.
$3,999.99
128 GB / 2 TB, Windows 11 Pro or Linux; Micro Center exclusive
- Memory
- 128 GB unified LPDDR5X
- Biggest model, comfortably
- AMD says up to 200B parameters
- Availability
- Micro Center, in-store pickup only
- Best for
- Developers who want a supported, Spark-sized box on x86 with Windows or Linux
What reviewers say
Written reviews
Video reviews
31:56Is STRIX Better than SPARK? Now Launching w/new Software: AMD's Ryzen AI Halo Developer WorkstationLevel1Techs · 31:56 · YouTube
17:42AMD Says 2 Ryzen AI Halos Can Run a 400B Model... I Tested ItAlex Ziskind · 17:42 · YouTube
34:10Making AMD's Ryzen AI Halo Do WorkCraft Computing · 34:10 · YouTube
From X
Glitch 81 ᯅ @deivid_techLocal AI tinkerer
“If you only compare the Ryzen AI Halo against the Nvidia DGX Spark at the RAM capacity, you don’t understand the importance of those 2 ports on the spark.…”
AI at AMD @AIatAMDMaker
“Run massive LLMs and AI workflows locally on AMD Ryzen AI Halo. Use it standalone, pair it with existing hardware, or plug it into a hybrid cloud workflow – the choice is yours.…”
Same chip, your favorite logo
GB10 partner boxes
ASUS, Dell, HP, Lenovo, Gigabyte and others
ASUS Ascent GX10, Dell Pro Precision with GB10, HP ZGX Nano G1n, Lenovo ThinkStation PGX and Gigabyte AI TOP ATOM all use the same GB10 chip and software as the DGX Spark. They differ in chassis, storage, support and price. MSI and Acer sell versions too. Right now Dell's 128 GB build, $5,903 with a 2 TB SSD, undercuts NVIDIA's own 128 GB Spark by about $1,000.
From $4,999 (64 GB)
64 GB from October 23; 128 GB from $5,903 (Dell, 2 TB), and most others run $7,000–$9,000
- Memory
- 64 GB or 128 GB unified (GB10)
- Biggest model, comfortably
- Same as DGX Spark
- Availability
- 128 GB patchy; 64 GB from October 23
- Best for
- Teams that want the Spark with a corporate warranty, bigger SSDs or a familiar vendor
What reviewers say
Written reviews
Video reviews
From X
HandwerkAiLab @HandwerkAiLabOwner
“…Tip for PGX owners: do EC/UEFI updates at the console via the DGX Dashboard, not over SSH. Users report SSH updates hang the reboot.…”
Damian Strobel @damian_89_Owner
“There they are - and yes those are ThinkStation PGX - its often cheaper BUT exactly the same Hardware and Software ;) Running them as 2 clusters.…”
The first RTX Spark desktop with a price
Surface RTX Spark Dev Box
Microsoft
Microsoft's compact desktop on NVIDIA's new Arm-based RTX Spark chip, the same platform as the RTX Spark laptops. It pairs 128 GB of unified memory with up to a petaflop of FP4 AI compute and ships with a developer setup of Windows 11 Pro, VS Code, GitHub Copilot and WSL. It costs $1,000 more than a 64 GB DGX Spark with twice the memory, but StorageReview notes it lacks the ConnectX-7 ports that let Sparks cluster.
$5,999.99
128 GB / 2 TB, Windows 11 Pro; one configuration, Microsoft.com (US) only
- Memory
- 128 GB unified (NVIDIA RTX Spark, 20-core CPU, 6,144 CUDA cores)
- Biggest model, comfortably
- Over 120B parameters (Microsoft)
- Availability
- Pre-order now; ships in November, US only
- Best for
- Developers who want CUDA and 128 GB on a Windows desk, with dev tools set up out of the box
What reviewers say Pre-release
Not shipping yet. These are launch coverage and posts from the company, not hands-on reviews.
Launch coverage
The repairable one
Framework Desktop
Framework
A 4.5-litre Mini-ITX workstation. Framework added a 192 GB Ryzen AI Max+ PRO 495 configuration on September 30 and recommends DeepSeek V4 Flash at IQ2 on the 128 GB box and DeepSeek V4.1 Flash at Q2 on the 192 GB box. The 495 model is Linux-only, and Framework warns memory prices will keep rising for six months.
From $6,799 (192 GB)
192 GB DIY Edition, pre-order, ships November; 128 GB mainboard alone is $3,149 (out of stock)
- Memory
- Up to 192 GB unified (Ryzen AI Max+ PRO 495, 273 GB/s)
- Biggest model, comfortably
- DeepSeek V4.1 Flash at Q2 on 192 GB (Framework's own pick)
- Availability
- 192 GB batch 1 on pre-order for November; 128 GB out of stock
- Best for
- Linux users who want standard Mini-ITX parts and a path to clustering
What reviewers say
Written reviews
Video reviews
From X
Sudo su @sudoingXOwner
“i'm very impressed with my framework desktop, amd's strix halo with 128gb of unified memory. it's the workstation where i make my content and where my agents cut my videos.…”
Cruz @ViC305Owner
“…Qwen3.8-Flash-Next-EXL3 on a Strix Halo GPU: ~17 → 40.0 tok/s mean, 46.3 peak. 🚀
That is on a 128 GB Framework Desktop.…”
코지베어 🐻 CozyBear @cozybearlogSkeptic
“did the math on the new Framework desktop before letting myself get excited. $6,799 buys 192GB and ports. it barely buys tokens.…”
Strix Halo, now with 192 GB
192 GB Ryzen AI Max+ PRO 495 boxes
GMKtec, Minisforum and others
AMD's refreshed Strix Halo chip lifts the memory ceiling from 128 GB to 192 GB with slightly faster memory (273 GB/s against 256 GB/s), so much bigger sparse models fit, but replies don't get much quicker. GMKtec's EVO-X5 Pro and Minisforum's MS-S1 MAX-P495 are the first mini PCs you can order with it; the 192 GB Framework Desktop uses the same chip. AMD's own 192 GB Ryzen AI Halo is headed to Micro Center, with no price yet.
From $6,799 (192 GB)
GMKtec EVO-X5 Pro, 192 GB / 2 TB (4 TB is $7,099); Minisforum MS-S1 MAX-P495 is $7,399 on a pre-order sale, list $9,249
- Memory
- 192 GB unified LPDDR5X (273 GB/s; up to 160 GB for the GPU)
- Biggest model, comfortably
- Up to 300B parameters (AMD)
- Availability
- EVO-X5 Pro shipping; MS-S1 MAX-P495 on pre-order, shipping from mid-October
- Best for
- Tinkerers who want room for 200B–300B mixture-of-experts models in a mini PC, on Windows or Linux
What reviewers say
Written reviews
The CUDA box
NVIDIA DGX Spark
NVIDIA
The GB10 Grace Blackwell chip with DGX OS, the full NVIDIA software stack and a 200 Gb ConnectX-7 port for pairing two units. Its 273 GB/s of memory bandwidth caps single-user generation speed; prompt processing and serving several users at once are where it pulls ahead, and “stuff just works,” as Tom's Hardware put it. It launched at $3,999 in October 2025; memory costs have since pushed the 128 GB model to $6,950.
$6,950 (128 GB)
NVIDIA's new list after the Oct 2 rise; 64 GB partner models from $4,999 on Oct 23
- Memory
- 64 GB or 128 GB unified (GB10, 273 GB/s)
- Biggest model, comfortably
- Up to 200B parameters on 128 GB; about 100B on 64 GB (NVIDIA)
- Availability
- 128 GB scarce; 64 GB on sale October 23 via partners
- Best for
- Developers who need CUDA, fine-tuning, or several agents served at once
What reviewers say
Written reviews
- ServeTheHome: NVIDIA DGX Spark Review The GB10 Machine is so Freaking Cool
- Tom's Hardware: Nvidia DGX Spark review: the GB10 Superchip powers a fast and fun AI toolbox that beats out AMD’s Ryzen AI Max+ 395
- Tom's Hardware: Nvidia introduces 64GB DGX Spark to throw local AI fans a lifeline amid the RAMpocalypse news
Video reviews
From X
Yume_X @yume_arasakiOwner (2× Spark)
“…I think I found the fastest dual DGX Spark recipe for Qwen3.8-Flash-Next. 138 tok/s structured, 133 code, 383 structured aggregate at four streams, and now it talks to Hermes.…”
Pratik Gandhi @pratikgBenchmark
“…What that means if you own a Spark, decode has gone from 17-18 tokens per second to 34.8.”
When a box isn't big enough
tinybox
tiny corp
Not a desk ornament. tiny corp, the company behind tinygrad, now sells a build-your-own tinybox: a 42-pound, water-cooled 32-core EPYC machine with a 1,700-watt power supply and room for four GPUs. Four Radeon 9070 XTs add $4,000, and NVIDIA RTX 6000 cards with 96 GB each are $19,000 apiece, up to two. It's paid by bank wire and is here as the upper bound of “AI at home.” The older red v2 and green v2 boxes are out of stock.
From $11,000
new build-your-own tinybox with 4× Radeon 9070 XT; $7,000 with no GPUs; red v2 ($12,000) and green v2 ($91,000) out of stock
- Memory
- 64 GB VRAM (4× Radeon 9070 XT) + 32 GB RAM; up to 192 GB VRAM with two NVIDIA RTX 6000 cards
- Biggest model, comfortably
- 70B class split across four GPUs (est.)
- Availability
- New tinybox made to order, ships in 1–8 weeks; red v2 and green v2 out of stock
- Best for
- Small labs and serious hobbyists with a 1,700 W circuit and room for a 42 lb box
What reviewers say
We couldn't find an independent written or video review of the v2 or new boxes.
From X
the tiny corp @__tinygrad__Maker
“It's tinygrad hosted Qwen3.8-27B in Pi Coding Agent playing Pokémon Red on a tinybox red by looking at the screen and pressing buttons.…”
Cloud vs box: what it really costs
A box costs money once; an API bills by the token. Here's what the same work costs in the cloud at the list prices on our model price board, blended 3 parts input to 1 part output, as of October 12, 2026. “Light” is 10 million tokens a month, about what a busy person chatting and coding through the day might use. “Heavy” is 60 million, an agent running around the clock.
| Cloud model | Blended $/1M | Light: 10M/mo | Heavy: 60M/mo | Months to repay a $3,499 box (heavy) |
|---|---|---|---|---|
| GPT-6 Luna | $0.200 | $2.00 | $12.00 | about 24 years |
| DeepSeek V4.1 Flash | $0.525 | $5.25 | $31.50 | about 9 years |
| Mistral Large 4 (Preview) | $2.065 | $20.65 | $123.90 | 28 months |
| Claude Sonnet 5.5 | $4.000 | $40.00 | $240.00 | 15 months |
| Claude Fable 5.1 | $20.000 | $200.00 | $1,200.00 | 3 months |
Payback ignores electricity. A box averaging 100 W around the clock uses about 72 kWh a month, roughly $13 at an assumed $0.18 per kWh, so subtract that from the monthly saving.
Questions people ask
What is a home AI computer?
A computer built or sold to run AI models on your own hardware instead of a cloud service. Some are appliances with models preinstalled, like Ghost Core and Oki Home. Others are general machines with lots of fast memory, like the Mac Studio, DGX Spark and AMD Strix Halo mini PCs, that you set up with tools such as Ollama, LM Studio or llama.cpp.
How much memory do I need to run a local AI model?
Roughly the model's file size plus room for context. A 27B model at 4-bit is about 18 GB (Framework lists Qwen3.8-27B at 17.6 GB), so 24 GB of GPU memory is comfortable. gpt-oss-120b in MXFP4 is about 63 GB, which needs a 96–128 GB unified-memory machine. Speed then depends on memory bandwidth.
Is a local AI box cheaper than paying for a cloud AI API?
Only at high volume against mid-priced models. At the Claude Sonnet 5.5 list price, 60 million tokens a month costs about $240, so a $3,499 box pays for itself in about 15 months before electricity. Against cheap tiers such as DeepSeek V4.1 Flash, the same usage costs about $32 a month and a box takes most of a decade to pay off.
Does running AI locally keep my data private?
The model and your files stay on the box, which removes the cloud provider from the loop. But features like remote access, connected accounts and optional cloud models can still send data out, as Oki's FAQ notes, so check what you switch on.
DGX Spark, Mac Studio or a Strix Halo mini PC?
Pick DGX Spark for CUDA, fine-tuning and serving several users or agents. Pick a Mac Studio for the fastest single-user replies. Pick a Strix Halo box such as the GMKtec EVO-X2 for the most memory per dollar on Windows or Linux.
Should I wait for prices to drop?
Not on current evidence. Memory shortages pushed the 128 GB DGX Spark from $3,999 at launch to $6,950, and Framework expects memory prices to keep rising for the next six months.