hacklas

Thanks. I agree, GGUF and upstream contributions are on my radar.

For a dense model this would be a limitation, but not all of a MoE model needs to be in memory, but the largest part of a MoE are the routed experts.

Some parts are needed to generated every single token and these really should fit in memory, but the router experts that are not neeed can rest on SSD and be read only if they are needed, so... you can run MoE models bigger than you memory, try the IQ2XXS.

It should work on your 64 GB after you enable SSD mode in DwarfStar (in MoEspesso it enables itself), while being slower, so... I am really hoping for good models between the 50-120 GB other than Laguna, there is a big gap right now unfortunately.

Worth mentioning why this is harder than it looks.

There is a different set of experts at every layer, and each layer has a small router that decides which ones to use.

The router needs to look at the state produced by the experts below it.

Drafted tokens from the MTP head can be used to predict which experts the first layer will want, but not beyond that. To know what layer 10 experts needs, you have to run layers 1-9 which means loading their experts.

So, yes, instead of a next-token drafter like MTP, you'd want something trained to predict the expert activation across all layers at once.

The Deepseek V3 is a model with 671 billion parameters, of which 37 billion are active.

Magistral Small is a 24 billion parameter model.

Pretty impressive in terms of efficiency for Mistral.

The size of the Magistral Medium is not publicly available, so it is difficult to compare efficiency there.