Can your machine run it?
The honest answer — fit, speed tier, assumptions stated.
36 GPUs · 14 systems · 40 models · 154 quants — sourced & stamped
Fit physics, not vibes
Weights + KV cache + overhead against your real memory pool — per quant, per context length, from measured file sizes.
Honest speed tiers
Interactive, usable, or painful — a tier we can stand behind, never a fake-precise tok/s number. Uncalibrated paths say “beta”.
Every assumption stated
Runtime, batch size, context, KV precision — printed on every result. If we don't know yet, the answer says so.
Popular checks
Run your own →Live verdicts at Q4_K_M · llama.cpp · 8K context — including the honest no's.
NVIDIA GeForce RTX 4090 × Llama-3.1-70B-Instruct
NVIDIA GeForce RTX 4090 × Qwen2.5-32B-Instruct
NVIDIA GeForce RTX 5090 × Qwen2.5-32B-Instruct
NVIDIA GeForce RTX 3090 × Qwen2.5-14B-Instruct
NVIDIA DGX Spark × gpt-oss-120b
Mac Studio M3 Ultra 96GB × Llama-3.1-70B-Instruct
GMKtec EVO-X2 Ryzen AI Max+ 395 128GB × gpt-oss-120b
AMD Radeon RX 7900 XTX × Qwen2.5-32B-Instruct
Intel Arc B580 × Llama-3.1-8B-Instruct
NVIDIA GeForce RTX 4090 × DeepSeek-R1
Mac mini M4 24GB × Qwen2.5-7B-Instruct
NVIDIA GeForce RTX 4090 × gpt-oss-20b
Three kinds of machine, one honest answer
Discrete GPUs
NVIDIA, AMD, and Intel cards — VRAM is the hard wall, bandwidth sets the speed. AMD/Intel verdicts ship with a beta label until calibrated.
Apple unified memory
Macs share one memory pool between CPU and GPU — huge models load, but bandwidth, not capacity, decides whether they're pleasant.
Mini-PC APUs
DGX Spark, Strix Halo boxes: 128GB pools at laptop-class bandwidth. Great for big MoE models, honest tier told straight — this class is new, so it's beta.
How the math works — in the open
The formulas, thresholds, and every calibration constant are published. Cite our results, or connect your AI assistant directly to the checker.