merge

This is a merge of pretrained gemma3 language models

Some circumstances made me remember what real low-end computer is, so idea of this merge was born.

Goal of this merge is to create uncensored, somewhat smart all-round model, small enough to run on really ancient hardware.

This model takes main good features from its components, it is not censored (to a reasonable degree, extremely stuff wasn't tested), creative, able to multilang AND have not bad uncensored vision.

This model tested on core2duo E8400 with 8gb ddr2 ram, partially offloaded on 2gb gt630. It was more an experiment, but speed was bearable. On normal hardware model is FAST.

Of course, it is 4B after all, so don't expect 32 or 24b performance, but for it's size it is really good. For vision i used mmproj from SicariusSicariiStuff/X-Ray_Alpha

tested both q5_k_m and q8, both stable, but i reccomend to use q8 if possible.