Mars_27B_V.1
This is a merge of pre-trained Gemma3 language models.
Goal of this merge was to create high quality roleplay model for complex cards and scenarios. ERP too.
The result i can call good. Model is smart, smarter than 24b mistral. Writing style is pleastant to read, instructions followed.
Gemma context attention is flawless, every small detail is remembered.
Cliches are present, though they are different from Mistral's.
Ru is supported, good enough for roleplay, good enough as assistant.
Image part was briefly tested, it works. I don't use that functionality so i can't tell more.
I had only problem with this model - it's too memory hungry, even on Q4_K_S and 8k context only half of layers fits it my 16gb vram.
Tested on T0.8, XTC off.