ConicCat/Gemma4-GarnetV3-31B

VISIONPricing:Input $0.48 / Cached $0.1 / Output $1.44Concurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ConicCat/Gemma4-GarnetV3-31B is a 31 billion parameter language model developed by ConicCat, fine-tuned from the Gemma 4 architecture. This model specializes in enhancing roleplay and writing performance, with a particular focus on prose quality and generating human-like characters. It leverages DPO training on a dataset comprising approximately one-third writing and two-thirds roleplay content. Its primary strength lies in generating high-quality narrative and interactive character-driven text.

Loading preview...

Model Overview

ConicCat/Gemma4-GarnetV3-31B is a 31 billion parameter model built upon the Gemma 4 architecture, developed by ConicCat. This iteration, GarnetV3, has been specifically fine-tuned using Direct Preference Optimization (DPO) to significantly improve its capabilities in roleplay and writing tasks. A core focus during its development was to enhance prose quality and the realism of generated characters, making it suitable for applications requiring nuanced and engaging textual outputs.

Key Capabilities

  • Enhanced Roleplay: Excels at generating dynamic and believable character interactions.
  • Improved Writing Quality: Produces text with a strong emphasis on prose quality and narrative flow.
  • Human-like Characters: Designed to create characters that feel authentic and engaging.

Training Details

The model was trained for 9 hours on a single A100 80GB GPU. Its training leveraged a unique dataset composition, approximately one-third dedicated to writing improvement and two-thirds to roleplay scenarios. The datasets used include:

  • ConicCat/Lamp_P_Preference: Focused on improving prose by comparing human-revised writing against AI-generated text.
  • ConicCat/Charcards_Delta_Qwen3_5V2: Utilized a delta tuning recipe for data generation, comparing Qwen3.5 27B against Qwen3.5 2B.
  • ConicCat/Charcards_Context_Distill_Gemma4_26BV2: Employed context distillation from Gemma 4 26B, using full context as preferred and missing context as rejected examples.

Recommended Usage

For optimal performance, users are recommended to utilize the Q4_K_M GGUF quantization with koboldcpp.

Finetunes
1 models