Will it run? Check one model on one GPU or Mac

Pick your hardware, a model and a quant. You get whether it fits in memory (with the numbers behind it), the measured speed if someone has run it, or a bandwidth estimate with the formula shown, and the next GPU up.

Hardware Memory Model Quant

How the answer is worked out

Fit. Memory needed = the GGUF file size from Hugging Face + a runtime and KV-cache allowance of 0.8 GB + 0.025 GB per billion parameters (capped at 3 GB), which covers about an 8K-token context. A discrete GPU can use its VRAM minus 0.5 GB; a Mac or APU can use 75% of its unified memory.

Speed. If a run of that model and quant on that exact hardware is in the data, you see the measured number and its source. Otherwise you see an estimate: effective bandwidth / max(1.2, GB read per token). Effective bandwidth comes from a measured run on the same chip when there is one, or spec bandwidth times the typical efficiency for that vendor. For mixture-of-experts models, GB read per token = file size × min(1, 2 × active share).

Next step up. If it does not fit, the smallest tracked desktop GPU with measured runs that holds it in VRAM. If it fits, the least powerful tracked desktop GPU that is at least 25% faster on it. Amazon links are affiliate links.

Measured pairs

The 40 model and hardware pairs with the most measured runs. Each page lists every run with its source.

Pair	Fastest quant	Best generation	Runs
Llama 2 7B on the AMD Radeon RX 5700 XT	Q4_0	74.4 measured	4
Llama 2 7B on the AMD Radeon RX 6900 XT	Q4_0	108 measured	4
Llama 2 7B on the AMD Radeon RX 7600 XT	Q4_0	56.1 measured	4
Llama 2 7B on the AMD Radeon RX 7800 XT	Q4_0	125 measured	4
Llama 2 7B on the AMD Radeon RX 7900 GRE	Q4_0	116 measured	4
Llama 2 7B on the AMD Radeon RX 7900 XT	Q4_0	123 measured	4
Llama 2 7B on the AMD Radeon RX 7900 XTX	Q4_0	191 measured	4
Qwen3.8 27B on the AMD Radeon RX 7900 XTX	UD-Q4_K_XL	41.0 measured	4
Llama 2 7B on the AMD Radeon RX 9060 XT	Q4_0	74.0 measured	4
Llama 2 7B on the AMD Radeon RX 9070	Q4_0	120 measured	4
Llama 2 7B on the AMD Radeon RX 9070 XT	Q4_0	137 measured	4
Llama 2 7B on the AMD Radeon RX Vega 64	Q4_0	67.7 measured	4
Llama 3.1 8B Instruct on the Apple M2 (10-core GPU)	Q4_K_M	14.7 measured	3
Llama 3.1 8B Instruct on the Apple M4 (10-core GPU)	Q4_K_M	16.8 measured	3
Llama 3.2 1B Instruct on the Apple M4 Max (40-core GPU)	Q4_K_M	180 measured	3
Qwen3 4B Instruct 2507 on the Apple M4 Max (40-core GPU)	MLX-4bit	135 measured	7
Qwen3 30B-A3B Instruct 2507 on the Apple M4 Max (40-core GPU)	MLX-4bit	113 measured	4
Llama 3.1 8B Instruct on the Apple M4 Max (40-core GPU)	Q4_K_M	52.9 measured	3
Llama 3.1 8B Instruct on the Apple M4 Pro (20-core GPU)	Q4_K_M	32.7 measured	3
Llama 3.1 8B Instruct on the Apple M5 (10-core GPU)	Q4_K_M	20.2 measured	3
Llama 2 7B on the NVIDIA GTX 1070 Ti	Q4_0	43.4 measured	4
Llama 2 7B on the NVIDIA GTX 1080 Ti	Q4_0	71.6 measured	4
Llama 2 7B on the NVIDIA RTX 2080 Ti	Q4_0	109 measured	4
Llama 3.1 8B Instruct on the NVIDIA RTX 3050	Q4_K_M	34.5 measured	3
Llama 2 7B on the NVIDIA RTX 3060	Q4_0	80.6 measured	4
Llama 2 7B on the NVIDIA RTX 3080	Q4_0	146 measured	4
Llama 2 7B on the NVIDIA RTX 3090	Q4_0	172 measured	4
Qwen3.8 27B on the NVIDIA RTX 3090	UD-Q2_K_XL	52.4 measured	8
Llama 2 7B on the NVIDIA RTX 4070 Ti SUPER	Q4_0	136 measured	4
Llama 2 7B on the NVIDIA RTX 4080 SUPER	Q4_0	150 measured	4
gpt-oss 20B on the NVIDIA RTX 4090	MXFP4	255 measured	3
Llama 2 7B on the NVIDIA RTX 4090	Q4_0	190 measured	4
Qwen3.8 27B on the NVIDIA RTX 4090	UD-Q4_K_XL	47.9 measured	7
Llama 2 7B on the NVIDIA RTX 5060 Ti	Q4_0	97.0 measured	4
Llama 2 7B on the NVIDIA RTX 5070 Ti	Q4_0	182 measured	4
gpt-oss 20B on the NVIDIA RTX 5080	MXFP4	229 measured	5
Llama 2 7B on the NVIDIA RTX 5080	Q4_0	192 measured	4
Llama 2 7B on the NVIDIA RTX 5090	Q4_0	300 measured	4
Qwen3.8 27B on the NVIDIA RTX 5090	Q4_K_M	76.9 measured	12
Llama 2 7B on the NVIDIA TITAN V	Q4_0	112 measured	4

Want the whole list for one card? Use What fits my card.