jtl11/INTELLECT-3-REAP-50-heretic
LoRA-style (fewer layers) vs PrimeIntellect/INTELLECT-3. jtl11/INTELLECT-3-REAP-50-heretic matches part of PrimeIntellect/INTELLECT-3 (9195 of its 17925 tensors; the rest of PrimeIntellect/INTELLECT-3 has no counterpart). On the shared part: Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of down_proj, gate_proj, o_proj, up_proj changed, in 45 of 46 layers; norms, biases and embeddings are untouched.
Tensors changed: Embeddings 0/1, Norms 0/93, MLP 1183/1331, Attention 30/322, lm_head 0/1
1 derivatives checked
- Akicou/INTELLECT-3-REAP-50-heretic-GGUF: Quantization