Ilya626/Cydonia_Vistral

TEXT GENERATIONPricing:Input $0.7 / Cached $0.04 / Output $1.16Concurrent Unit Cost:2Model Size:24BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 26, 2025Architecture:Transformer0.0K Featherless Exclusive Cold

Ilya626/Cydonia_Vistral is a 24 billion parameter language model created by Ilya626, developed using the SLERP merge method to combine Vistral-24B-Instruct and Cydonia-24B-v4.2.0. This model integrates the strengths of its constituent models, with a gradient SLERP profile that blends Vistral's characteristics at lower layers and Cydonia's at higher layers. It is designed for general language tasks, leveraging a combined vocabulary and supporting Llama 3 or Mistral Tekken chat templates.

Loading preview...

Model Overview

Ilya626/Cydonia_Vistral is a 24 billion parameter language model resulting from a sophisticated merge of two pre-trained models: Vistral-24B-Instruct and Cydonia-24B-v4.2.0. This model was constructed using the SLERP (Spherical Linear Interpolation) merge method, which allows for a nuanced combination of the source models' weights.

Merge Details

The merge configuration employed a gradient SLERP profile, specifically designed to blend the characteristics of Vistral and Cydonia across different transformer layers. The t parameter in the SLERP configuration was varied for self_attn and mlp filters, creating a smooth transition where lower layers lean more towards Vistral and higher layers incorporate more of Cydonia. This approach aims to harness the distinct strengths of each base model.

Key Features

  • 24 Billion Parameters: A substantial model size for robust language understanding and generation.
  • SLERP Merge Method: Utilizes a advanced merging technique for balanced integration of source models.
  • Combined Vocabulary: Employs a union tokenizer source to ensure comprehensive vocabulary coverage from both Vistral and Cydonia.
  • Flexible Chat Templates: Supports popular chat formats, including Llama 3 and Mistral Tekken, for broad compatibility in conversational AI applications.

Intended Use Cases

This model is suitable for a variety of general-purpose language tasks where a blend of capabilities from its constituent models would be beneficial. Its architecture suggests potential for applications requiring nuanced understanding and generation, leveraging the combined knowledge and stylistic properties of Vistral and Cydonia.

Finetunes
2 models