How deep can your machine go? 🛗
A one-page field guide to running open-weight LLMs locally, on a Mac, a shared-memory mini PC, or a
gaming GPU.
Live page: https://niznik-dev.github.io/local-llm-elevator/
Models are drawn as floors in an elevator shaft: small models sit near the surface and the giants
are far below. Every machine's elevator only goes so deep, and its stop marks where it runs out of
memory. The shaft's built-in stops are reference machines (Macs and mini PCs in Unified mode, graphics cards in VRAM mode); add your own in two steps:
- Pick a memory type: Unified or GPU VRAM (the ? explains the difference).
- Pick a preset or type your own amount: 🍎 Macs or a mini PC (Strix Halo, DGX Spark) for unified, 🎮 gaming cards for VRAM. Press
Enter or ↓ to ride down to your stop; Clear removes it.
| Mode | Enter | Usable estimate |
|---|
| Unified | Total RAM on Apple Silicon or another shared-memory machine (AMD Strix Halo, NVIDIA DGX Spark) | ~67% (≤36GB) or ~75%, leaving room for the OS and a few apps |
| GPU VRAM | A dedicated GPU's VRAM, summed across cards (presets: RX 9060 XT 8GB, RTX 5070 12GB, RX 9070 XT 16GB, RTX 3090 24GB, RTX 5090 32GB, with a 1× / 2× / 4× card count; counts a card isn't usually combined at are greyed out) | ~90%, after driver and display overhead |
VRAM mode assumes the whole model fits on the GPU. Partial offload to system RAM (llama.cpp layer
splitting, MoE experts on CPU) is out of scope.
At ground level the elevator waits in the lobby of a cutaway RobCo-ish tower, with a pneumatic tube
up to the rider's second-floor office. Below the grass the shaft digs through soil, an aquifer,
stratified rock, and finally the mantle, where the car starts to shake.
Once you've set a stop, the elevator parks there and everything deeper dims, so you can see at a
glance what's out of reach. The rider's shirt and your stop take the vendor's color when you pick a
GPU preset (NVIDIA green, AMD red) and stay orange otherwise.
A money-green hardware value box floats above the floor tracker: the US price of the preset you
picked (times the card count), or, for a typed amount, an estimate from the cheapest preset's $/GB.
GPU picks add a small line estimating the rest of the machine (PC, workstation, or server, by card
count). Hover it for where the number comes from.
What's on each floor
| Field | Meaning |
|---|
| Q4 / Q8 memory | Approximate weight footprint at 4-bit and 8-bit quantization |
| Models | Notable open-weight releases from the past year at that size, linked to Hugging Face |
| Fit pills | The smallest reference machine that holds the Q4 / Q8 weights. Unified mode names Macs and mini PCs; VRAM mode names card setups with their total VRAM. Setups with equal room take turns. Past a single Mac, clustered 512GB Studios (Thunderbolt 5 RDMA, macOS 26.2+) appear as dashed theoretical pills |
| Cheapest option | The least expensive setup that holds the Q4 weights at default limits, within the current memory mode (a card isn't compared to a whole computer). GPU setups rank by cards plus host machine and show both, e.g. $2,400 + ~$2,000 for the PC; Mac clusters are marked theoretical |
| 🏢 Data center | How many 80GB H100s or ~180GB B200s hold the 8-bit (FP8) weights, at 90% usable per card (minimums, not deployment-shaped counts) |
| Comparable Claude model | The Claude tier you'd reach for on the same kind of task (not a claim of equal performance) |
| Comparable ChatGPT model | The same idea for OpenAI's API tiers (GPT-6 Luna, GPT-6.1 Terra, GPT-6.1 Sol, GPT-6 Astra) |
| Popularity | A rough three-step signal from OpenRouter token usage and Hugging Face downloads |
| Gem | Decoration. It's a mine shaft, so gems get roughly pricier the deeper you go (quartz to red diamond, then antimatter). They say nothing about the model |
Reference stops are shown for 32GB and 64GB MacBook Pros (💻) and 128 / 256 / 512GB Mac Studios (🤍 🩶 🖤), plus 128GB Strix Halo (❤️) and DGX Spark (💚) mini PCs, which take turns with the 128GB Studio in the fit pills. In VRAM mode those stops swap out for one per GPU preset, plus a stop for every multi-card count the selector allows (2× RX 9070 XT; 2× and 4× RTX 3090 and 5090), and one 8× RTX 5090 🔥 stop that's deliberately left out of the picker. A
solid box is the last floor that fits; a dashed box is the recommended stop if you want to keep
other apps open.
Rules of thumb
- Weights: about 0.6GB per billion parameters at 4-bit, about 1.1GB at 8-bit. Leave room for
context and the OS on top of that.
- Mixture-of-experts: an MoE model needs memory for all its parameters but runs at the speed of
the few active per token.
- Shared-memory caps: on a Mac, the GPU can use roughly two-thirds to three-quarters of unified
memory by default; raise it with
sudo sysctl iogpu.wired_limit_mb=<MB> (resets on reboot). Strix
Halo sets the GPU's share in the BIOS or AMD's Variable Graphics Memory; DGX Spark has no fixed split.
- The VRAM wall: a discrete card has no cap to raise. What's on the card is what you get, minus a
little for the driver and display.
Usage
No build step and no dependencies. Open index.html in a browser:
open index.html
The page loads fonts from Google Fonts and otherwise works offline.
Updating the data
All model and device data lives in a JSON block inside index.html:
| Key | Holds |
|---|
"levels" | Each floor: size, Q4/Q8 memory, models, comparable Claude and ChatGPT models, gem |
"devices" | Reference Mac stops: usable / raised GPU memory, price, and an optional cluster list of theoretical Thunderbolt 5 cluster sizes (pills and cheapest line only, no stops) |
"minis" | Unified mini PCs (Strix Halo, DGX Spark): usable memory, icon, price |
"gpus" | VRAM presets: vendor (nvidia or amd) for the dot and shirt color, price per card, max cards (with a why tooltip for greyed-out counts), rigStops, and an optional showoff count |
"hosts" | Rough cost of the machine around 1, 2, 4, or 8 GPUs (PC, workstation, server) |
"datacenter" | Cards for the 🏢 line, with a url for each name's link |
To add a model, append it to the matching level's models array; to change the memory budget
assumptions, edit appsReserveGB and contextReserveGB. Prices are US prices as of November 2026
(street prices for GPUs); add "priceEst": true to mark one as a rough estimate. Bump "updated"
when you do.
Caveats
Memory figures are approximate and vary by quant format and runtime. Check each model card before you
download 400GB of weights.
License
MIT © 2026 Mattie Niznik
