Nexesenex/Llama_3.x_70b_Hexagon_Purple_V3
Nexesenex/Llama_3.x_70b_Hexagon_Purple_V3 is a 70 billion parameter language model from the Llama 3.x family, developed by Nexesenex. This model is a merge of several specialized Llama 3.x variants, created using the Model Stock method with Nexesenex/Llama_3.x_70b_SmarTricks_V1.01 as its base. It incorporates components like Gutenberg-Doppel, HighPriestess, Electra, DarkHorse, and FLDx2-Tess3, aiming for a balanced performance profile across various tasks. This V3 iteration refines the V2 version by adjusting specific merged components to subtly alter its characteristics.
Loading preview...
Model Overview
Nexesenex/Llama_3.x_70b_Hexagon_Purple_V3 is a 70 billion parameter model built upon the Llama 3.x architecture, developed by Nexesenex. It is a composite model, created through a "Model Stock" merge using mergekit, with Nexesenex/Llama_3.x_70b_SmarTricks_V1.01 serving as its foundational base model.
Key Components and Refinements
This V3 iteration is a refinement of its V2 predecessor, incorporating several specialized Llama 3.x variants:
- Base Model:
Nexesenex/Llama_3.x_70b_SmarTricks_V1.01 - Merged Models:
nbeerbower/Llama3.1-Gutenberg-Doppel-70BNexesenex/Llama_3.1_70b_HighPriestess_R1_V1(upgraded with Lumitron Lorablated)Steelskull/L3.3-Electra-R1-70bNexesenex/Llama_3.3_70b_DarkHorse(replaces DoppelGangerR1 from V2)Nexesenex/Llama_3.1_70b_FLDx2-Tess3_abliterated_fusion_norm(Tess merged with Hitachi FLDx2)
Differentiators
This model's primary distinction lies in its sophisticated merging strategy, combining diverse Llama 3.x models to achieve a specific performance balance. The V3 update specifically adjusts the blend of components like DarkHorse and Smartricks to subtly modify its characteristics compared to V2, aiming for a nuanced improvement. Users of V2 may find this version quite similar, with minor differences in output behavior due to the component adjustments.
Intended Use
Given its nature as a merged model from various Llama 3.x derivatives, Hexagon Purple V3 is likely suitable for general-purpose language generation tasks where a robust 70B parameter model is desired. Its specific blend of components suggests a focus on balancing different aspects of language understanding and generation, though precise strengths would require empirical testing.