⚠️ Attention: It works with ChatML format and Mistral V7-Tekken (no thinking output) i didn't test the Thinking.

Tested using Ik_llama.cpp as the backend and Sillytavern as the frontend

🌟 Overview

Rei-Pulse-24B is a merge built on top of Morax-24B-v1.

The model uses the DARE TIES method

🧩 Models Merged

This model results from a merge between:

• DarkArtsForge/Morax-24B-v1 (Base)
• OddTheGreat/Circuitry_24B_V.3
• Delta-Vector/Rei-24B-KTO
• TheDrummer/Precog-24B-v1
• TheDrummer/Rivermind-24B-v1

🛠️ Merge Details

Method: DARE TIES

The merge was performed using mergekit with the following parameters:

• Base Model: Morax-24B-v1
• Dtype: bfloat16
• Density: 0.5
• Weight: 0.2

⚙️ Configuration

base_model: DarkArtsForge/Morax-24B-v1
dtype: bfloat16
merge_method: dare_ties
modules:
  default:
    slices:
    - sources:
      - layer_range: [0, 40]
        model: DarkArtsForge/Morax-24B-v1
        parameters:
          density: 0.5
          weight: 0.2
      - layer_range: [0, 40]
        model: TheDrummer/Rivermind-24B-v1
        parameters:
          density: 0.5
          weight: 0.2
      - layer_range: [0, 40]
        model: OddTheGreat/Circuitry_24B_V.3
        parameters:
          density: 0.5
          weight: 0.2
      - layer_range: [0, 40]
        model: Delta-Vector/Rei-24B-KTO
        parameters:
          density: 0.5
          weight: 0.2
      - layer_range: [0, 40]
        model: TheDrummer/Precog-24B-v1
        parameters:
          density: 0.5
          weight: 0.2
parameters:
  lambda: 0.75
  normalize: True
tokenizer:
  source: base

Static GGUF quant Imatrix GGUF quant

EXL3 quant

✨ Acknowledgements

Thanks to the authors of the original models for their incredible work:

• DarkArtsForge for Morax-24B-v1
• OddTheGreat for Circuitry_24B_V.3
• Delta-Vector for Rei-24B-KTO
• TheDrummer for Precog-24B-v1 and Rivermind-24B-v1