1 Trillion Parameter
1T models like GPT, full QWEN, Llama, Ling-1T, Kimi-K2-Chat/Base all will be runnable on consumer hardware soon (Now if you want to pioneer it).
How, lets break it down.
The biggest contributing factor making this possible is GPUDirect Storage
The biggest limiting factor is 63 GB/s PCI-E 5.0 bandwidth and trying to saturate that bandwidth using consumer NVME which currently are at around 7.0GBS (Double that for Enterprise Drives)
Third limiting factor, any model loaded in this fashion must have offloading, if it requires the model be fully loaded then it is out.
Other considerations losses do to poor tensor shape management. LBA blocks are 4096KB and the the quantized tensor needs to ideally be in a factor of 1024
Power Consumption is a fraction of GPU consumption in this assume 25-30W per NvME (100W realistically in a GRAID setup)
Even with all that said current available tech such as motherboards with 3 pci-e 5.0 NVME slots and with a high end GPU (24-32GB) you could run the NVFP4 1T model in around 22-38 seconds per query.
Fun Fact: A 4096-byte LBA block holds 8192 NVFP4 numbers, and the number of possible data combinations 2^32,768
2^32,768 seconds, a number so large it exceeds the age of the universe by more than 9,800 orders of magnitude.
Maybe we should encrypt in NVFP4 LBA block combinations makes SHA-256 and 512 look like a joke.