YFC-112358/Qwen3.8-27B-Della-Carnice-Ostrich-Salience-Glimmer-v5c

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 1, 2026Architecture:Transformer Featherless Exclusive Cold

YFC-112358/Qwen3.8-27B-Della-Carnice-Ostrich-Salience-Glimmer-v5c is a 27 billion parameter language model based on Qwen/Qwen3.8-27B, created by YFC-112358. This model is a single-stage 'della_linear' merge of four distinct source models, designed to combine their strengths. It utilizes a specific merging recipe with weighted contributions and density parameters for each source, focusing on a unique tensor-level fusion approach. The model is optimized for nuanced integration of diverse model characteristics through its 'della_linear' merge method, making it suitable for applications requiring a blend of capabilities from its constituent models.

Loading preview...

Model Overview

This model, YFC-112358/Qwen3.8-27B-Della-Carnice-Ostrich-Salience-Glimmer-v5c, is a 27 billion parameter language model built upon the Qwen/Qwen3.8-27B base. It employs a single-stage della_linear merge method, integrating four distinct source models: kai-os/Carnice-V3, etemiz/Ostrich-27B-Qwen3.8-260815, vectionlabs/Salience-27B-R5, and YFC-112358/Qwen3.8-27B-TM-Gemma4-Glimmer-v3.

Key Characteristics

  • Fusion Method: Utilizes a della_linear merge, which performs a weighted linear combination of tensors from the base and source models. This method does not involve 'sign voting', ensuring all source contributions are directly integrated.
  • Parameter Blending: Each source model contributes with specific weight, density, and epsilon parameters, influencing its effective presence in the final model. The normalize: true setting means nominal weights determine relative proportions, not absolute strength.
  • Tensor-Level Operations: The merging process involves combining 866 tensors and directly transferring 333 tensors, totaling 1199 tensors. Magnitude pruning is applied within row blocks, and surviving terms are rescaled for unbiased representation.
  • Precision: Calculations are performed in fp32, with the final model saved in bf16 format.
  • Tokenizer & Config: The tokenizer and configuration are inherited directly from Qwen/Qwen3.8-27B, ensuring consistency with the base model's tokenization scheme.

Use Cases

This model is particularly well-suited for developers looking to leverage a composite model that combines the strengths of multiple fine-tuned models based on Qwen3.8-27B. Its unique della_linear merging approach offers a distinct way to blend model characteristics without traditional consensus mechanisms, potentially leading to novel performance profiles for various NLP tasks.

Finetunes
1 models