maldv/Gemma-4-31B-Isometry-Fabled-Persona

VISIONPricing:Input $0.48 / Cached $0.1 / Output $1.44Concurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 11, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The maldv/Gemma-4-31B-Isometry-Fabled-Persona is a 31 billion parameter Gemma 4-based model developed by Praxis Maldevide, fine-tuned for reasoning, agentic behavior, and conversational capabilities. It features a unique merge strategy combining various LoRAs and NVFP4 deltas to balance strong reasoning with a distinct creative and roleplay voice. This model excels at complex instruction following and provides a more present conversational style than pure reasoning models, while maintaining a 32768 token context length.

Loading preview...

Model Overview

maldv/Gemma-4-31B-Isometry-Fabled-Persona is a 31 billion parameter model built upon google/gemma-4-31B-it, developed by Praxis Maldevide. It is engineered as a reasoning-first, agentic, and conversational model, distinguishing itself through a sophisticated merge process that carefully integrates various finetunes. This approach ensures a distinct voice for creative and roleplay tasks without compromising its core reasoning abilities.

Key Capabilities & Differentiators

  • Advanced Merge Strategy: Utilizes a unique method to extract and balance different types of updates (LoRAs, NVFP4 deltas) from various sources, preserving signal integrity and behavioral nuances.
  • Reasoning & Agentic Behavior: Incorporates strong reasoning capabilities from sources like TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2 and agentic Fable behavior from TeichAI/Gemma-4-31B-Fable-5-Agent-Distill-LoRA.
  • Balanced Persona: Achieves a blend of instruction discipline, Fable-style planning, and a more engaging conversational voice, suitable for creative and character-driven responses.
  • Refusal Control: Includes a sparse low-rank LoRA from trohrbaugh/gemma-4-31b-it-heretic-ara to intentionally modify refusal behavior.
  • Performance: Limited screening shows balanced performance, giving up some peak GSM8K reasoning scores for substantially improved strict instruction following compared to individual source models.

Recommended Use Cases

  • Complex Reasoning Tasks: Ideal for applications requiring strong analytical and problem-solving skills.
  • Agentic Workflows: Suitable for scenarios where the model needs to exhibit planning and agent-like behaviors.
  • Conversational AI: Excellent for chatbots and interactive applications that benefit from a distinct, creative, and engaging persona.
  • Creative Writing & Roleplay: Provides atmospheric creative and character responses without being exclusively an RP-first model.

Finetunes
1 models