Hardware

Every hardware sheet answers the question the spec sheet dodges: how much memory is really left once the system takes its share, and which of the 18 measured models fit in what remains — with the context each one can actually hold. Two lists per machine, the ones that fit and the ones that do not, so you can see what not to buy.

MachineMemory left for the modelBandwidthModels that fit at 8kand at 32kLargest model it takes
PC with an RTX 4060 8 GB7.20 GiB272 GB/s
derived, not published
5 / 183 / 18qwen3-8b 4.68 GiB
PC with an RTX 3060 12 GB11.20 GiB360 GB/s
derived, not published
6 / 186 / 18gemma4-12b 6.63 GiB
PC with an RTX 4070 12 GB11.20 GiB504 GB/s
derived, not published
6 / 186 / 18gemma4-12b 6.63 GiB
Mac with M4 and 16 GB unified memory13.00 GiB120 GB/s
vendor declared
7 / 187 / 18gpt-oss-20b 10.83 GiB
Office laptop, CPU only, 16 GB13.00 GiB83 GB/s
estimated
7 / 187 / 18gpt-oss-20b 10.83 GiB
Mac with M4 Pro and 24 GB21.00 GiB273 GB/s
vendor declared
13 / 1811 / 18qwen3-coder-30b 17.28 GiB
Workstation with an RTX 4090 24 GB23.20 GiB1008 GB/s
vendor declared
14 / 1813 / 18qwen3.6-35b-a3b 20.61 GiB
Mini-PC with an NPU and 32 GB LPDDR5X29.00 GiB120 GB/s
estimated
14 / 1814 / 18qwen3.6-35b-a3b 20.61 GiB
Server with 2× RTX 3090 (48 GB)47.20 GiB936 GB/s
vendor declared
14 / 1814 / 18qwen3.6-35b-a3b 20.61 GiB
Mac with M4 Max and 64 GB61.00 GiB546 GB/s
vendor declared
15 / 1815 / 18gpt-oss-120b 58.46 GiB

Fitting means the model file, its context memory and the runtime headroom all fit in what is left after the system takes its share, at Q4_K_M or the nearest compression the author publishes. Bandwidth carries its provenance in the cell: five of these ten figures are not published by the manufacturer anywhere, and we say so rather than passing them off as specifications.