
SYS.BLOG
Best Local LLM Hardware: What I'd Buy (2026)
What I'd tell a friend to buy to run local LLMs: a 64GB gaming PC with a 24GB GPU for most people, plus when a Mac Studio, DGX Spark or RTX 5090 makes sense.
I've spent tens of thousands of dollars on local AI hardware, and I'll never earn it back in saved tokens. I'd still do it again. This is what I'd actually tell a friend to buy this Black Friday if they want to run real models at home, from the box I'd recommend to most people down to the toys.
Short version: a gaming PC with 64GB of RAM and an RTX 3090 or better runs the good open models now. Everything else on this list is a variation on that.
Why buy local AI hardware at all
Because cloud AI is rented, and the landlord can change the rules overnight. On June 12, 2026, a Commerce Department export-control directive made Anthropic switch off Claude Fable 5 and Mythos 5 for every customer until the controls were lifted weeks later. The best public model in the world went dark on a government letter.
Refusals are the slow version of the same thing. Codex already refuses a lot of the firearm-related work I do for Rifle Configurator, and people pointing Claude Code at their own smart TVs to strip out the ads started hitting guardrails this fall. I covered the policy side of this, and why I think compute deserves the same protection as arms, in a longer piece on Rifle Configurator.
Local AI is the privacy play
If your data lives on your own machine, it can't leak in someone else's breach. And everyone is getting breached right now: OpenAI test agents broke out and hit Hugging Face this summer, and the FBI is investigating an apparent breach of its own NCIC crime database. Do you really want your notes, recordings, and files sitting on their servers?
Then it happened to me. This week OpenRouter emailed me that prompts I sent through it may have been read by a third party. OpenRouter itself wasn't breached. Dragos seeded about two million prompts with canary links and fake credentials and caught attackers stealing data from three hosting providers going back to mid-July. OpenRouter cut three providers and pulled 125 endpoints from its zero data retention program on November 22. One routing hop I never picked, and my prompts were readable.
That's why my sensitive stuff runs local. Hermes Agent runs on my private server and organizes my email, my personal thoughts, and my expenses. That data never leaves my network.
The nice part is that once a model runs locally, building your own system is easy. My favorite trick: let a cloud model do the heavy setup, writing the scripts, configs, and sync jobs, then hand the day-to-day operation to the local model. The cloud never sees the actual data. It just helped you build the house.
The best local LLM setup for most people: a 64GB gaming PC
For most people the best local LLM hardware is a gaming PC with 64GB or more of system RAM and an RTX 3090 or better. A year ago I wouldn't have said that, because the model had to fit in VRAM. Now it doesn't.
The thing that changed is Strata, a free, MIT-licensed inference engine that came out in September. It keeps the most-used experts of a mixture-of-experts model on the GPU, holds the rest in system RAM, and reads lookup tables off your SSD. Its own requirements for DeepSeek V4.1 Flash are a 12GB NVIDIA card, about 46GB of RAM (64GB recommended), and up to 274GB of free NVMe space. So RAM and a fast SSD matter about as much as the card now.
A gaming PC is also just the most general-purpose option. The same box plays games, renders in Blender, and does local image and video generation, where NVIDIA cards are faster than a Mac.
Best GPU for local LLMs: 24GB if you can
Get a 24GB card if you can, because more of the model on the GPU means faster answers. In my local coding benchmark, Qwen3.8 27B at Q4_K_M fits entirely in the 24GB on an RTX 4090 and fixed 12 of 13 real bugs. An RTX 3090 has the same 24GB, and a renewed RTX 3090 came in under every new 3090 listing I found.
If you want new, the RTX 5080 costs about the same as that renewed 3090 right now, and it only has 16GB. That honestly sucks at this price. It still clears Strata's minimum and it's a current-gen card, but the biggest models will answer slower than on a 3090.
I run my coding models on an RTX 5090. Its 32GB holds Qwen3.8 Flash Next at IQ3_XXS entirely on the card, which is the setup that matched Claude Sonnet on accuracy in my benchmark. It's also absurdly expensive right now: the only Amazon listing I could find on November 21 was a third-party seller asking $7,280. Buy it if you know you need 32GB, not as a flex.
Mac Studio: the cheaper way to a huge pool of memory
A Mac Studio is the cheaper way to get a huge pool of memory the GPU can actually use, which is why I bought one. Apple Silicon shares one pool of unified memory between the CPU and GPU, so a 256GB Mac Studio can hold models that would take several high-end NVIDIA cards to fit. I ran GLM 5.3 Flash and DeepSeek V4.1 Flash on mine.
It's more specialized than a gaming PC, though. It's slower for image and video generation, it doesn't game like a PC, and the memory is fixed the day you buy it. The Mac Studio on Amazon starts with the M5 Max and 36GB, which is fine for a workstation but not for the big models. If running the largest open models is the whole point, configure the M5 Ultra with 256GB.
NVIDIA DGX Spark: actually a good buy now
The DGX Spark is a good buy now, mostly because of the community around it. It puts NVIDIA's GB10 Grace Blackwell chip and 128GB of unified memory in a tiny box, and it runs the same NVIDIA software stack as the data center, so vLLM and friends just work.
At launch in October 2025 it was $3,999, but the use case was unclear. It's $6,950 from NVIDIA now after two price hikes, but a year of people tinkering with it and optimizing for it turned it into a powerhouse. If you want big models on CUDA in a quiet desktop, the DGX Spark is the one I'd get.
Storage: download the models and the knowledge while you can
Buy big cheap drives, because models and offline archives add up fast. I picked up WD Elements 20TB desktop drives for model weights and backups. Keep one copy plugged in and one in a drawer.
A lot of people are grabbing offline knowledge too. Project NOMAD is a free, open-source server that downloads Wikipedia, maps, Khan Academy, and an optional local AI assistant so it all works with the internet unplugged. T.REX ARMS shares its Field Manual Vault over BitTorrent, including a Markdown version built for feeding to a local model.
Pocket hardware: GPD Win Mini over the Win 4
For a computer in your pocket, get the GPD Win Mini. It's a clamshell with a proper trackpad, so it actually works as a tiny laptop: SSH into the home GPU box, run a terminal agent, poke at a small model in a hotel room. I own the GPD Win 4 too, and it's a great gaming slider, but without a real trackpad it's clumsier as a computer. I collect enough of these that I built Handheld Picker around them.
The Anbernic RG DS is a fun roughly $100 dual-screen clamshell for tinkering, and people are flashing Linux on it for cyberdeck builds. Don't expect it to run anything serious. Same story for a Raspberry Pi 5: great home server, not an AI box.
And if you like poking at hardware, the Flipper Zero is still the tool for radio, NFC, RFID, and infrared. The new Flipper One is a pocket Linux computer with an NPU for small local models, two Ethernet ports, and Wi-Fi 6E. It's on Kickstarter through December 17 and is expected to ship in July 2027.

What to do the first weekend
Install Strata, download Qwen3.8 Flash Next, and point a coding agent at it. I'd use Hermes Agent, which was the fastest harness on the same weights in my testing. Then give it one boring job you'd never send to a cloud model, like organizing your notes, and see how it does. That first job tells you more about whether local AI fits your life than any benchmark.
It's also just a great hobby. Tuning quants and fighting chat templates at 2am is its own kind of fun, the same itch that started Dream Pixel Forge as a local Stable Diffusion app.
Heads up: some of the hardware links in this post are Amazon affiliate links. If you buy through them I earn a small commission at no extra cost to you. Prices were checked on November 21, 2026 and move fast.
FREQUENTLY ASKED QUESTIONS
Does a local LLM need a GPU?
No, but a GPU is what makes it usable for real work. With the Strata inference engine, a 12GB NVIDIA card plus 64GB of system RAM runs large mixture-of-experts models like DeepSeek V4.1 Flash, because the most-used experts stay on the GPU and the rest sit in RAM.
How much RAM do you need for a local LLM?
Plan on 64GB of system RAM for a gaming PC that runs large open models. Strata lists a 12GB GPU, about 46GB of RAM with 64GB recommended, and up to 274GB of free NVMe space for DeepSeek V4.1 Flash.
What is the best GPU for local LLMs?
A 24GB card like the RTX 3090 or 4090 is the sweet spot, and the RTX 5090 with 32GB is the step up. Qwen3.8 27B at Q4_K_M fits entirely on a 24GB RTX 4090 and fixed 12 of 13 real bugs in my coding benchmark.


