AuriAetherwiing/G4-31B-Musica-v1

Hugging Face
VISIONConcurrency Cost:2Model Size:31BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 27, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

AuriAetherwiing/G4-31B-Musica-v1 is a 31 billion parameter Gemma-4 based model, fine-tuned for roleplay, story generation, and conversational tasks. It is the second model in the Musica series, offering improved instruction following and steerability over its base model. This model excels in creative text generation and maintains strong reasoning capabilities, with non-existent refusals.

Loading preview...

Overview

AuriAetherwiing/G4-31B-Musica-v1 is a 31 billion parameter model based on Gemma-4, specifically fine-tuned for roleplay, story generation, and conversational AI. It represents the second iteration in the Musica series, building upon the TQ3.5-27B-Musica-v1. This version is noted for being a significant upgrade over the stock Gemma model in its target domains.

Key Capabilities

  • Enhanced Instruction Following: Demonstrates strong steerability and adherence to instructions, making it highly adaptable for various prompts.
  • Creative Text Generation: Excels in generating engaging roleplay scenarios, compelling stories, and natural conversations.
  • Reasoning: Supports both reasoning and non-reasoning tasks, with a mechanism to encourage DeepSeek-esque reasoning by prefilling with <|channel>thought\n Okay, let's see.
  • Refusal-Free: Users report non-existent refusals, ensuring a more consistent and cooperative generation experience.
  • Improved Diversity: Offers better swipe diversity compared to the base Gemma model.

Training Details

The model was trained using a r64a64 LoRA configuration over one epoch, taking approximately 35 hours on 2xRTX Pro 6000 Blackwell GPUs. The training utilized a pretokenized dataset with an 8192 sequence length, focusing on last-turn training to address Gemma-specific challenges. The training run was sponsored by ArliAI.