ajtaltarabukin2022/dare1

Hugging Face
TEXT GENERATIONConcurrency Cost:2Model Size:32BQuant:FP8Ctx Length:32kPublished:Apr 19, 2026Architecture:Transformer Warm

The ajtaltarabukin2022/dare1 is a 32 billion parameter language model created by ajtaltarabukin2022, merged using the DARE TIES method with Qwen/Qwen3-32B as its base. This model leverages a specific merging technique to combine pre-trained language models, aiming to enhance performance or introduce new capabilities. It is suitable for applications requiring a large language model derived from advanced merging strategies.

Loading preview...

Model Overview

The ajtaltarabukin2022/dare1 is a 32 billion parameter language model developed by ajtaltarabukin2022. It was created using the mergekit tool, specifically employing the DARE TIES merge method. The base model for this merge was Qwen/Qwen3-32B.

Merge Details

The DARE TIES (Disentangled Activation REgularization for Task-Independent Ensemble Selection) method was utilized to combine the base model with another component, ./vector_merge1. This technique involves a specific configuration to blend the weights of the constituent models, with a density and mask method value of 0.65 and a rescale factor of 1.0. The merge process targeted layers from 0 to 64 for both the base and the merged component.

Potential Use Cases

  • Research into model merging techniques: This model serves as an example of a DARE TIES merge, useful for studying its effects.
  • Applications requiring a Qwen3-32B derivative: Users seeking a model based on Qwen3-32B with potentially altered characteristics due to the merge.
  • Experimentation with merged model performance: Suitable for evaluating how DARE TIES merging impacts a large language model's capabilities.