Sakalti/ultiima-108B
Partial fine-tune (fewer layers) vs Sakalti/ultiima-72B. Sakalti/ultiima-72B matches part of Sakalti/ultiima-108B (963 of its 1443 tensors; the rest of Sakalti/ultiima-108B has no counterpart). On the shared part: Partial change: 48% of the weights changed (attention, mlp, norms), in 40 of 80 layers.
Tensors changed: lm_head 0/1, Embeddings 0/1, Norms 80/161, MLP 120/240, Attention 280/560
1 derivatives checked
- mradermacher/ultiima-108B-i1-GGUF: Quantization